Introduction to Algorithms for Data Mining and

Machine Learning 1st edition - eBook PDF pdf

download

https://ebooksecure.com/download/introduction-to-algorithms-for-
data-mining-and-machine-learning-ebook-pdf/

Download more ebook from https://ebooksecure.com

We believe these products will be a great fit for you. Click
the link to download now, or visit ebooksecure.com
to discover even more!

Machine Learning for Biometrics: Concepts, Algorithms

and Applications (Cognitive Data Science in Sustainable
Computing) 1st Edition - eBook PDF

https://ebooksecure.com/download/machine-learning-for-biometrics-
concepts-algorithms-and-applications-cognitive-data-science-in-
sustainable-computing-ebook-pdf/

BIG DATA ANALYTICS: Introduction to Hadoop, Spark, and

Machine-Learning 1st Edition - eBook PDF

https://ebooksecure.com/download/big-data-analytics-introduction-
to-hadoop-spark-and-machine-learning-ebook-pdf/

(eBook PDF) Introduction to Machine Learning with

Python: A Guide for Data Scientists

http://ebooksecure.com/product/ebook-pdf-introduction-to-machine-
learning-with-python-a-guide-for-data-scientists/

(eBook PDF) Introduction to Business Data Mining 1st

Edition

http://ebooksecure.com/product/ebook-pdf-introduction-to-
business-data-mining-1st-edition/
(eBook PDF) Machine Learning Refined: Foundations,
Algorithms, and Applications

http://ebooksecure.com/product/ebook-pdf-machine-learning-
refined-foundations-algorithms-and-applications/

(eBook PDF) Introduction to Data Mining, Global Edition

2nd Edition

http://ebooksecure.com/product/ebook-pdf-introduction-to-data-
mining-global-edition-2nd-edition/

Big Data Mining for Climate Change 1st edition - eBook

PDF

https://ebooksecure.com/download/big-data-mining-for-climate-
change-ebook-pdf/

(eBook PDF) Introduction to Data Mining 2nd Edition by

Pang-Ning Tan

http://ebooksecure.com/product/ebook-pdf-introduction-to-data-
mining-2nd-edition-by-pang-ning-tan/

Machine Learning for Planetary Science 1st Edition -

eBook PDF

https://ebooksecure.com/download/machine-learning-for-planetary-
science-ebook-pdf/
 Xin-She Yang

Introduction to
Algorithms for Data Mining
and Machine Learning
Introduction to Algorithms for Data Mining and
Machine Learning
This page intentionally left blank
Introduction to
Algorithms for Data
Mining and Machine
Learning

Xin-She Yang
Middlesex University
School of Science and Technology
London, United Kingdom
Academic Press is an imprint of Elsevier
125 London Wall, London EC2Y 5AS, United Kingdom
525 B Street, Suite 1650, San Diego, CA 92101, United States
50 Hampshire Street, 5th Floor, Cambridge, MA 02139, United States
The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, United Kingdom
Copyright © 2019 Elsevier Inc. All rights reserved.
No part of this publication may be reproduced or transmitted in any form or by any means, electronic or
mechanical, including photocopying, recording, or any information storage and retrieval system, without
permission in writing from the publisher. Details on how to seek permission, further information about the
Publisher’s permissions policies and our arrangements with organizations such as the Copyright Clearance Center
and the Copyright Licensing Agency, can be found at our website: www.elsevier.com/permissions.
This book and the individual contributions contained in it are protected under copyright by the Publisher (other
than as may be noted herein).
Notices
Knowledge and best practice in this field are constantly changing. As new research and experience broaden our
understanding, changes in research methods, professional practices, or medical treatment may become necessary.
Practitioners and researchers must always rely on their own experience and knowledge in evaluating and using
any information, methods, compounds, or experiments described herein. In using such information or methods
they should be mindful of their own safety and the safety of others, including parties for whom they have a
professional responsibility.
To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors, assume any liability
for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or
from any use or operation of any methods, products, instructions, or ideas contained in the material herein.

Library of Congress Cataloging-in-Publication Data

A catalog record for this book is available from the Library of Congress

British Library Cataloguing-in-Publication Data

A catalogue record for this book is available from the British Library

ISBN: 978-0-12-817216-2

For information on all Academic Press publications

visit our website at https://www.elsevier.com/books-and-journals

Publisher: Candice Janco

Acquisition Editor: J. Scott Bentley
Editorial Project Manager: Michael Lutz
Production Project Manager: Nilesh Kumar Shah
Designer: Miles Hitchen
Typeset by VTeX
Contents

About the author ix

Preface xi
Acknowledgments xiii

1 Introduction to optimization 1
1.1 Algorithms 1
1.1.1 Essence of an algorithm 1
1.1.2 Issues with algorithms 3
1.1.3 Types of algorithms 3
1.2 Optimization 4
1.2.1 A simple example 4
1.2.2 General formulation of optimization 7
1.2.3 Feasible solution 9
1.2.4 Optimality criteria 10
1.3 Unconstrained optimization 10
1.3.1 Univariate functions 11
1.3.2 Multivariate functions 12
1.4 Nonlinear constrained optimization 14
1.4.1 Penalty method 15
1.4.2 Lagrange multipliers 16
1.4.3 Karush–Kuhn–Tucker conditions 17
1.5 Notes on software 18

2 Mathematical foundations 19
2.1 Convexity 20
2.1.1 Linear and affine functions 20
2.1.2 Convex functions 21
2.1.3 Mathematical operations on convex functions 22
2.2 Computational complexity 22
2.2.1 Time and space complexity 24
2.2.2 Complexity of algorithms 25
2.3 Norms and regularization 26
2.3.1 Norms 26
2.3.2 Regularization 28
2.4 Probability distributions 29
2.4.1 Random variables 29
2.4.2 Probability distributions 30
vi Contents

2.4.3 Conditional probability and Bayesian rule 32

2.4.4 Gaussian process 34
2.5 Bayesian network and Markov models 35
2.6 Monte Carlo sampling 36
2.6.1 Markov chain Monte Carlo 37
2.6.2 Metropolis–Hastings algorithm 37
2.6.3 Gibbs sampler 39
2.7 Entropy, cross entropy, and KL divergence 39
2.7.1 Entropy and cross entropy 39
2.7.2 DL divergence 40
2.8 Fuzzy rules 41
2.9 Data mining and machine learning 42
2.9.1 Data mining 42
2.9.2 Machine learning 42
2.10 Notes on software 42

3 Optimization algorithms 45
3.1 Gradient-based methods 45
3.1.1 Newton’s method 45
3.1.2 Newton’s method for multivariate functions 47
3.1.3 Line search 48
3.2 Variants of gradient-based methods 49
3.2.1 Stochastic gradient descent 50
3.2.2 Subgradient method 51
3.2.3 Conjugate gradient method 52
3.3 Optimizers in deep learning 53
3.4 Gradient-free methods 56
3.5 Evolutionary algorithms and swarm intelligence 58
3.5.1 Genetic algorithm 58
3.5.2 Differential evolution 60
3.5.3 Particle swarm optimization 61
3.5.4 Bat algorithm 61
3.5.5 Firefly algorithm 62
3.5.6 Cuckoo search 62
3.5.7 Flower pollination algorithm 63
3.6 Notes on software 64

4 Data fitting and regression 67

4.1 Sample mean and variance 67
4.2 Regression analysis 69
4.2.1 Maximum likelihood 69
4.2.2 Liner regression 70
4.2.3 Linearization 75
4.2.4 Generalized linear regression 77
4.2.5 Goodness of fit 80
Contents vii

4.3 Nonlinear least squares 81

4.3.1 Gauss–Newton algorithm 82
4.3.2 Levenberg–Marquardt algorithm 85
4.3.3 Weighted least squares 85
4.4 Overfitting and information criteria 86
4.5 Regularization and Lasso method 88
4.6 Notes on software 90

5 Logistic regression, PCA, LDA, and ICA 91

5.1 Logistic regression 91
5.2 Softmax regression 96
5.3 Principal component analysis 96
5.4 Linear discriminant analysis 101
5.5 Singular value decomposition 104
5.6 Independent component analysis 105
5.7 Notes on software 108

6 Data mining techniques 109

6.1 Introduction 110
6.1.1 Types of data 110
6.1.2 Distance metric 110
6.2 Hierarchy clustering 111
6.3 k-Nearest-neighbor algorithm 112
6.4 k-Means algorithm 113
6.5 Decision trees and random forests 115
6.5.1 Decision tree algorithm 115
6.5.2 ID3 algorithm and C4.5 classifier 116
6.5.3 Random forest 120
6.6 Bayesian classifiers 121
6.6.1 Naive Bayesian classifier 121
6.6.2 Bayesian networks 123
6.7 Data mining for big data 124
6.7.1 Characteristics of big data 124
6.7.2 Statistical nature of big data 125
6.7.3 Mining big data 125
6.8 Notes on software 127

7 Support vector machine and regression 129

7.1 Statistical learning theory 129
7.2 Linear support vector machine 130
7.3 Kernel functions and nonlinear SVM 133
7.4 Support vector regression 135
7.5 Notes on software 137
viii Contents

8 Neural networks and deep learning 139

8.1 Learning 139
8.2 Artificial neural networks 140
8.2.1 Neuron models 140
8.2.2 Activation models 141
8.2.3 Artificial neural networks 143
8.3 Back propagation algorithm 146
8.4 Loss functions in ANN 147
8.5 Optimizers and choice of optimizers 149
8.6 Network architecture 149
8.7 Deep learning 151
8.7.1 Convolutional neural networks 151
8.7.2 Restricted Boltzmann machine 157
8.7.3 Deep neural nets 158
8.7.4 Trends in deep learning 159
8.8 Tuning of hyperparameters 160
8.9 Notes on software 161

Bibliography 163

Index 171
About the author

Xin-She Yang obtained his PhD in Applied Mathematics from the University of Ox-
ford. He then worked at Cambridge University and National Physical Laboratory (UK)
as a Senior Research Scientist. Now he is Reader at Middlesex University London, and
an elected Bye-Fellow at Cambridge University.
He is also the IEEE Computer Intelligence Society (CIS) Chair for the Task Force
on Business Intelligence and Knowledge Management, Director of the International
Consortium for Optimization and Modelling in Science and Industry (iCOMSI), and
an Editor of Springer’s Book Series Springer Tracts in Nature-Inspired Computing
(STNIC).
With more than 20 years of research and teaching experience, he has authored
10 books and edited more than 15 books. He published more than 200 research pa-
pers in international peer-reviewed journals and conference proceedings with more
than 36 800 citations. He has been on the prestigious lists of Clarivate Analytics and
Web of Science highly cited researchers in 2016, 2017, and 2018. He serves on the
Editorial Boards of many international journals including International Journal of
Bio-Inspired Computation, Elsevier’s Journal of Computational Science (JoCS), In-
ternational Journal of Parallel, Emergent and Distributed Systems, and International
Journal of Computer Mathematics. He is also the Editor-in-Chief of the International
Journal of Mathematical Modelling and Numerical Optimisation.
This page intentionally left blank
Preface

Both data mining and machine learning are becoming popular subjects for university
courses and industrial applications. This popularity is partly driven by the Internet and
social media because they generate a huge amount of data every day, and the under-
standing of such big data requires sophisticated data mining techniques. In addition,
many applications such as facial recognition and robotics have extensively used ma-
chine learning algorithms, leading to the increasing popularity of artificial intelligence.
From a more general perspective, both data mining and machine learning are closely
related to optimization. After all, in many applications, we have to minimize costs,
errors, energy consumption, and environment impact and to maximize sustainabil-
ity, productivity, and efficiency. Many problems in data mining and machine learning
are usually formulated as optimization problems so that they can be solved by opti-
mization algorithms. Therefore, optimization techniques are closely related to many
techniques in data mining and machine learning.
Courses on data mining, machine learning, and optimization are often compulsory
for students, studying computer science, management science, engineering design, op-
erations research, data science, finance, and economics. All students have to develop
a certain level of data modeling skills so that they can process and interpret data for
classification, clustering, curve-fitting, and predictions. They should also be familiar
with machine learning techniques that are closely related to data mining so as to carry
out problem solving in many real-world applications. This book provides an introduc-
tion to all the major topics for such courses, covering the essential ideas of all key
algorithms and techniques for data mining, machine learning, and optimization.
Though there are over a dozen good books on such topics, most of these books are
either too specialized with specific readership or too lengthy (often over 500 pages).
This book fills in the gap with a compact and concise approach by focusing on the key
concepts, algorithms, and techniques at an introductory level. The main approach of
this book is informal, theorem-free, and practical. By using an informal approach all
fundamental topics required for data mining and machine learning are covered, and
the readers can gain such basic knowledge of all important algorithms with a focus
on their key ideas, without worrying about any tedious, rigorous mathematical proofs.
In addition, the practical approach provides about 30 worked examples in this book
so that the readers can see how each step of the algorithms and techniques works.
Thus, the readers can build their understanding and confidence gradually and in a
step-by-step manner. Furthermore, with the minimal requirements of basic high school
mathematics and some basic calculus, such an informal and practical style can also
enable the readers to learn the contents by self-study and at their own pace.
This book is suitable for undergraduates and graduates to rapidly develop all the
fundamental knowledge of data mining, machine learning, and optimization. It can
xii Preface

also be used by students and researchers as a reference to review and refresh their
knowledge in data mining, machine learning, optimization, computer science, and data
science.

Xin-She Yang
January 2019 in London
Acknowledgments

I would like to thank all my students and colleagues who have given valuable feedback
and comments on some of the contents and examples of this book. I also would like to
thank my editors, J. Scott Bentley and Michael Lutz, and the staff at Elsevier for their
professionalism. Last but not least, I thank my family for all the help and support.

Xin-She Yang
January 2019
This page intentionally left blank
Introduction to optimization
Contents
1.1 Algorithms
1 1
1.1.1 Essence of an algorithm 1
1.1.2 Issues with algorithms 3
1.1.3 Types of algorithms 3
1.2 Optimization 4
1.2.1 A simple example 4
1.2.2 General formulation of optimization 7
1.2.3 Feasible solution 9
1.2.4 Optimality criteria 10
1.3 Unconstrained optimization 10
1.3.1 Univariate functions 11
1.3.2 Multivariate functions 12
1.4 Nonlinear constrained optimization 14
1.4.1 Penalty method 15
1.4.2 Lagrange multipliers 16
1.4.3 Karush–Kuhn–Tucker conditions 17
1.5 Notes on software 18

This book introduces the most fundamentals and algorithms related to optimization,
data mining, and machine learning. The main requirement is some understanding of
high-school mathematics and basic calculus; however, we will review and introduce
some of the mathematical foundations in the first two chapters.

1.1 Algorithms
An algorithm is an iterative, step-by-step procedure for computation. The detailed
procedure can be a simple description, an equation, or a series of descriptions in
combination with equations. Finding the roots of a polynomial, checking if a natu-
ral number is a prime number, and generating random numbers are all algorithms.

1.1.1 Essence of an algorithm

In essence, an algorithm can be written as an iterative equation or a set of iterative
equations. For example, to find a square root of a > 0, we can use the following
iterative equation:
1 a
xk+1 = xk + , (1.1)
2 xk
where k is the iteration counter (k = 0, 1, 2, . . . ) starting with a random guess x0 = 1.
Introduction to Algorithms for Data Mining and Machine Learning. https://doi.org/10.1016/B978-0-12-817216-2.00008-9
Copyright © 2019 Elsevier Inc. All rights reserved.
2 Introduction to Algorithms for Data Mining and Machine Learning

Example 1
As an example, if x0 = 1 and a = 4, then we have

1 4
x1 = (1 + ) = 2.5. (1.2)
2 1

Similarly, we have

1 4 1 4
x2 = (2.5 + ) = 2.05, x3 = (2.05 + ) ≈ 2.0061, (1.3)
2 2.5 2 2.05
x4 ≈ 2.00000927, (1.4)
√
which is very close to the true value of 4 = 2. The accuracy of this iterative formula or algorithm
is high because it achieves the accuracy of five decimal places after four iterations.

The convergence is very quick if we start from different initial values such as
x0 = 10 and even x0 = 100. However, for an obvious reason, we cannot start with
x0 = 0 due to division by
√zero.
Find the root of x = a is equivalent to solving the equation

f (x) = x 2 − a = 0, (1.5)

which is again equivalent to finding the roots of a polynomial f (x). We know that
Newton’s root-finding algorithm can be written as

f (xk )
xk+1 = xk − , (1.6)
f  (xk )

where f  (x) is the first derivative or gradient of f (x). In this case, we have
f  (x) = 2x. Thus, Newton’s formula becomes

(xk2 − a)
xk+1 = xk − , (1.7)
2xk

which can be written as

xk a 1 a
xk+1 = (xk − )+ = xk + ). (1.8)
2 2xk 2 xk

This is exactly what we have in Eq. (1.1).

Newton’s method has rigorous mathematical foundations, which has a guaranteed
convergence under certain conditions. However, in general, Eq. (1.6) is more general,
and the gradient information f  (x) is needed. In addition, for the formula to be valid,
we must have f  (x) = 0.
Introduction to optimization 3

1.1.2 Issues with algorithms

The advantage of the algorithm given in Eq. (1.1) is that√it converges very quickly.
However, careful readers may have asked: we know that 4 = ±2, how can we find
the other root −2 in addition to +2?
Even if we use different initial value x0 = 10 or x0 = 0.5, we can only reach x∗ = 2,
not −2.
What happens if we start with x0 < 0? From x0 = −1, we have
1 4 1 4
x1 = (−1 + ) = −2.5, x 2 = (−2.5 + ) = −2.05, (1.9)
2 −1 2 −2.5
x3 ≈ −2.0061, x4 ≈ −2.00000927, (1.10)
which is approaching −2 very quickly. If we start from x0 = −10 or x0 = −0.5, then
we can always get x∗ = −2, not +2.
This highlights a key issue here: the final solution seems to depend on the initial
starting point for this algorithm, which is true for many algorithms.
Now the relevant question is: how do we know where to start to get a particular
solution? The general short answer is “we do not know”. Thus, some knowledge of
the problem under consideration or an educated guess may be useful to find the final
solution.
In fact, most algorithms may depend on the initial configuration, and such algo-
rithms are often carrying out search moves locally. Thus, this type of algorithm is
often referred to as local search. A good algorithm should be able to “forget” its initial
configuration though such algorithms may not exist at all for most types of problems.
What we need in general is the global search, which attempts to find final solutions
that are less sensitive to the initial starting point(s).
Another important issue in our discussions is that the gradient information f  (x) is
necessary for some algorithms such as Newton’s method given in Eq. (1.6). This poses
certain requirements on the smoothness of the function f (x). For example, we know
that |x| is not differentiable at x = 0. Thus, we cannot directly use Newton’s method
to find the roots of f (x) = |x|x 2 − a = 0 for a > 0. Some modifications are needed.
There are other issues related to algorithms such as the setting of parameters, the
slow rate of convergence, condition numbers, and iteration structures. All these make
algorithm designs and usage somehow challenging, and we will discuss these issues
in more detail later in this book.

1.1.3 Types of algorithms

An algorithm can only do a specific computation task (at most a class of computational
tasks), and no algorithms can do all the tasks. Thus, algorithms can be classified due
to their purposes. An algorithm to find roots of a polynomial belongs to root-finding
algorithms, whereas an algorithm for ranking a set of numbers belongs to sorting
algorithms. There are many classes of algorithms for different purposes. Even for the
same purpose such as sorting, there are many different algorithms such as the merge
sort, bubble sort, quicksort, and others.
4 Introduction to Algorithms for Data Mining and Machine Learning

We can also categorize algorithms in terms of their characteristics. The root-finding

algorithms we just introduced are deterministic algorithms because the final solutions
are exactly the same if we start from the same initial guess. We obtain the same set of
solutions every time we run the algorithm. On the other hand, we may introduce some
randomization into the algorithm, for example, using purely random initial points.
Every time we run the algorithm, we use a new random initial guess. In this case, the
algorithm can have some nondeterministic nature, and such algorithms are referred
to as stochastic.√Sometimes, using randomness may be advantageous. For example, in
the example of 4 = ±2 using Eq. (1.1), random initial values (both positive and neg-
ative) can allow the algorithm to find both roots. In fact, a major trend in the modern
metaheuristics is using some randomization to suit different purposes.
For algorithms to be introduced in this book, we are mainly concerned with al-
gorithms for data mining, optimization, and machine learning. We use a relatively
unified approach to link algorithms in data mining and machine learning to algorithms
for optimization.

1.2 Optimization

Optimization is everywhere, from engineering design to business planning. After all,

time and resources are limited, and optimal use of such valuable resources is crucial.
In addition, designs of products have to maximize the performance, sustainability, and
energy efficiency and to minimize the costs. Therefore, optimization is important for
many applications.

1.2.1 A simple example

Let us start with a very simple example to design a container with volume capacity
V0 = 10 m3 . As the main cost is related to the cost of materials, the main aim is to
minimize the total surface area S.
The first thing we have to decide is the shape of the container (cylinder, cubic,
sphere or ellipsoid, or more complex geometry). For simplicity, let us start with a
cylindrical shape with radius r and height h (see Fig. 1.1).
The total surface area of a cylinder is

S = 2(πr 2 ) + 2πrh, (1.11)

and the volume is

V = πr 2 h. (1.12)

There are only two design variables r and h and one objective function S to be min-
imized. Obviously, if there is no capacity constraint, then we can choose not to build
the container, and then the cost of materials is zero for r = 0 and h = 0. However,
Introduction to optimization 5

Figure 1.1 Design of a cylindric container.

the constraint requirement means that we have to build a container with fixed volume
V0 = πr 2 h = 10 m3 . Therefore, this optimization problem can be written as

minimize S = 2πr 2 + 2πrh, (1.13)

subject to the equality constraint

πr 2 h = V0 = 10. (1.14)

To solve this problem, we can first try to use the equality constraint to reduce the
number of design variables by solving h. So we have
V0
h= . (1.15)
πr 2
Substituting it into (1.13), we get

S = 2πr 2 + 2πrh
V0 2V0
= 2πr 2 + 2πr 2 = 2πr 2 + . (1.16)
πr r
This is a univariate function. From basic calculus we know that the minimum or max-
imum can occur at the stationary point, where the first derivative is zero, that is,
dS 2V0
= 4πr − 2 = 0, (1.17)
dr r
which gives

V0 3 V0
r3 = , or r = . (1.18)
2π 2π
Thus, the height is

h V0 /(πr 2 ) V0
= = 3 = 2. (1.19)
r r πr
6 Introduction to Algorithms for Data Mining and Machine Learning

This means that the height is twice the radius: h = 2r. Thus, the minimum surface is

S∗ = 2πr 2 + 2πrh = 2πr 2 + 2πr(2r) = 6πr 2

 V 2/3 6π
0 2/3
= 6π =√3
V0 . (1.20)
2π 4π 2

For V0 = 10, we have

 
3 V0 3 10
r= = ≈ 1.1675, h = 2r = 2.335,
(2π) 2π

and the total surface area

S∗ = 2πr 2 + 2πrh ≈ 25.69.

It is worth pointing out that this optimal solution is based on the assumption or re-
quirement to design a cylindrical container. If we decide to use a sphere with radius R,
we know that its volume and surface area is
4π 3
V0 = R , S = 4πR 2 . (1.21)
3
We can solve R directly

3V0 3 3V0
R =
3
, or R = , (1.22)
4π 4π
which gives the surface area
 3V 2/3 √
0 4π 3 9 2/3
S = 4π =√ 3
V0 . (1.23)
4π 16π 2
√3 √ √ 3
Since 6π/ 4π 2 ≈ 5.5358 and 4π 3 9/ 16π 2 ≈ 4.83598, we have S < S∗ , that is, the
surface area of a sphere is smaller than the minimum surface area of a cylinder with
the same volume. In fact, for the same V0 = 10, we have
√
4π 3 9 2/3
S(sphere) = √ 3
V0 ≈ 22.47, (1.24)
16π 2
which is smaller than S∗ = 25.69 for a cylinder.
This highlights the importance of the choice of design type (here in terms of shape)
before we can do any truly useful optimization. Obviously, there are many other fac-
tors that can influence the choice of design, including the manufacturability of the
design, stability of the structure, ease of installation, space availability, and so on. For
a container, in most applications, a cylinder may be much easier to produce than a
sphere, and thus the overall cost may be lower in practice. Though there are so many
factors to be considered in engineering design, for the purpose of optimization, here
we will only focus on the improvement and optimization of a design with well-posed
mathematical formulations.
Introduction to optimization 7

1.2.2 General formulation of optimization

Whatever the real-world applications may be, it is usually possible to formulate an
optimization problem in a generic form [49,53,160]. All optimization problems with
explicit objectives can in general be expressed as a nonlinearly constrained optimiza-
tion problem

maximize/minimize f (x), x = (x1 , x2 , . . . , xD )T ∈ RD ,

subject to φj (x) = 0 (j = 1, 2, . . . , M),
ψk (x) ≤ 0 (k = 1, . . . , N), (1.25)

where f (x), φj (x), and ψk (x) are scalar functions of the design vector x. Here the
components xi of x = (x1 , . . . , xD )T are called design or decision variables, and they
can be either continuous, discrete, or a mixture of these two. The vector x is often
called the decision vector, which varies in a D-dimensional space RD .
It is worth pointing out that we use a column vector here for x (thus with trans-
pose T ). We can also use a row vector x = (x1 , . . . , xD ) and the results will be the
same. Different textbooks may use slightly different formulations. Once we are aware
of such minor variations, it should cause no difficulty or confusion.
In addition, the function f (x) is called the objective function or cost function,
φj (x) are constraints in terms of M equalities, and ψk (x) are constraints written as
N inequalities. So there are M + N constraints in total. The optimization problem
formulated here is a nonlinear constrained problem. Here the inequalities ψk (x) ≤ 0
are written as “less than”, and they can also be written as “greater than” via a simple
transformation by multiplying both sides by −1.
The space spanned by the decision variables is called the search space RD , whereas
the space formed by the values of the objective function is called the objective or
response space, and sometimes the landscape. The optimization problem essentially
maps the domain RD or the space of decision variables into the solution space R (or
the real axis in general).
The objective function f (x) can be either linear or nonlinear. If the constraints φj
and ψk are all linear, it becomes a linearly constrained problem. Furthermore, when
φj , ψk , and the objective function f (x) are all linear, then it becomes a linear pro-
gramming problem [35]. If the objective is at most quadratic with linear constraints,
then it is called a quadratic programming problem. If all the values of the decision
variables can be only integers, then this type of linear programming is called integer
programming or integer linear programming.
On the other hand, if no constraints are specified and thus xi can take any values
in the real axis (or any integers), then the optimization problem is referred to as an
unconstrained optimization problem.
As a very simple example of optimization problems without any constraints, we
discuss the search of the maxima or minima of a univariate function.
8 Introduction to Algorithms for Data Mining and Machine Learning

2
Figure 1.2 A simple multimodal function f (x) = x 2 e−x .

Example 2
For example, to find the maximum of a univariate function f (x)

f (x) = x 2 e−x ,
2
−∞ < x < ∞, (1.26)

is a simple unconstrained problem, whereas the following problem is a simple constrained mini-
mization problem:

f (x1 , x2 ) = x12 + x1 x2 + x22 , (x1 , x2 ) ∈ R2 , (1.27)

subject to

x1 ≥ 1, x2 − 2 = 0. (1.28)

It is worth pointing out that the objectives are explicitly known in all the optimiza-
tion problems to be discussed in this book. However, in reality, it is often difficult to
quantify what we want to achieve, but we still try to optimize certain things such as the
degree of enjoyment or service quality on holiday. In other cases, it may be impossible
to write the objective function in any explicit form mathematically.
From basic calculus we know that, for a given curve described by f (x), its gradient
f  (x) describes the rate of change. When f  (x) = 0, the curve has a horizontal tangent
at that particular point. This means that it becomes a point of special interest. In fact,
the maximum or minimum of a curve occurs at
f  (x∗ ) = 0, (1.29)

which is a critical condition or stationary condition. The solution x∗ to this equation

corresponds to a stationary point, and there may be multiple stationary points for a
given curve.
To see if it is a maximum or minimum at x = x∗ , we have to use the information of
its second derivative f  (x). In fact, f  (x∗ ) > 0 corresponds to a minimum, whereas
f  (x∗ ) < 0 corresponds to a maximum. Let us see a concrete example.

Example 3
To find the minimum of f (x) = x 2 e−x (see Fig. 1.2), we have the stationary condition
2

f  (x) = 0 or

f  (x) = 2x × e−x + x 2 × (−2x)e−x = 2(x − x 3 )e−x = 0.

2 2 2
Introduction to optimization 9

Figure 1.3 (a) Feasible domain with nonlinear inequality constraints ψ1 (x) and ψ2 (x) (left) and linear
inequality constraint ψ3 (x). (b) An example with an objective of f (x) = x 2 subject to x ≥ 2 (right).

As e−x > 0, we have

2

x(1 − x 2 ) = 0, or x = 0 and x = ±1.

The second derivative is given by

f  (x) = 2e−x (1 − 5x 2 + 2x 4 ),
2

which is an even function with respect to x.

So at x = ±1, f  (±1) = 2[1 − 5(±1)2 + 2(±1)4 ]e−(±1) = −4e−1 < 0. Thus, there are
2

two maxima that occur at x∗ = ±1 with fmax = e−1 . At x = 0, we have f  (0) = 2 > 0, thus
the minimum of f (x) occurs at x∗ = 0 with fmin (0) = 0.

Whatever the objective is, we have to evaluate it many times. In most cases, the
evaluations of the objective functions consume a substantial amount of computational
power (which costs money) and design time. Any efficient algorithm that can reduce
the number of objective evaluations saves both time and money.
In mathematical programming, there are many important concepts, and we will
first introduce a few related concepts: feasible solutions, optimality criteria, the strong
local optimum, and weak local optimum.

1.2.3 Feasible solution

A point x that satisfies all the constraints is called a feasible point and thus is a feasible
solution to the problem. The set of all feasible points is called the feasible region (see
Fig. 1.3).
For example, we know that the domain f (x) = x 2 consists of all real numbers. If
we want to minimize f (x) without any constraint, all solutions such as x = −1, x = 1,
and x = 0 are feasible. In fact, the feasible region is the whole real axis. Obviously,
x = 0 corresponds to f (0) = 0 as the true minimum.
However, if we want to find the minimum of f (x) = x 2 subject to x ≥ 2, then it
becomes a constrained optimization problem. The points such as x = 1 and x = 0 are
no longer feasible because they do not satisfy x ≥ 2. In this case the feasible solutions
are all the points that satisfy x ≥ 2. So x = 2, x = 100, and x = 108 are all feasible. It
is obvious that the minimum occurs at x = 2 with f (2) = 22 = 4, that is, the optimal
solution for this problem occurs at the boundary point x = 2 (see Fig. 1.3).
10 Introduction to Algorithms for Data Mining and Machine Learning

Figure 1.4 Local optima, weak optima, and global optimality.

1.2.4 Optimality criteria

A point x ∗ is called a strong local maximum of the nonlinearly constrained op-
timization problem if f (x) is defined in a δ-neighborhood N (x ∗ , δ) and satisfies
f (x ∗ ) > f (u) for u ∈ N (x ∗ , δ), where δ > 0 and u = x ∗ . If x ∗ is not a strong lo-
cal maximum, then the inclusion of equality in the condition f (x ∗ ) ≥ f (u) for all
u ∈ N (x ∗ , δ) defines the point x ∗ as a weak local maximum (see Fig. 1.4). The local
minima can be defined in a similar manner when > and ≥ are replaced by < and ≤,
respectively.
Fig. 1.4 shows various local maxima and minima. Point A is a strong local max-
imum, whereas point B is a weak local maximum because there are many (in fact,
infinite) different values of x that will lead to the same value of f (x ∗ ). Point D is the
global maximum, and point E is the global minimum. In addition, point F is a strong
local minimum. However, point C is a strong local minimum, but it has a discontinuity
in f  (x ∗ ). So the stationary condition for this point f  (x ∗ ) = 0 is not valid. We will
not deal with these types of minima or maxima in detail.
As we briefly mentioned before, for a smooth curve f (x), optimal solutions usu-
ally occur at stationary points where f  (x) = 0. This is not always the case because
optimal solutions can also occur at the boundary, as we have seen in the previous ex-
ample of minimizing f (x) = x 2 subject to x ≥ 2. In our present discussion, we will
assume that both f (x) and f  (x) are always continuous or f (x) is everywhere twice
continuously differentiable. Obviously, the information of f  (x) is not sufficient to
determine whether a stationary point is a local maximum or minimum. Thus, higher-
order derivatives such as f  (x) are needed, but we do not make any assumption at this
stage. We will further discuss this in detail in the next section.

1.3 Unconstrained optimization

Optimization problems can be classified as either unconstrained or constrained. Un-

constrained optimization problems can in turn be subdivided into univariate and mul-
tivariate problems.
Introduction to optimization 11

1.3.1 Univariate functions

The simplest optimization problem without any constraints is probably the search for
the maxima or minima of a univariate function f (x). For unconstrained optimization
problems, the optimality occurs at the critical points given by the stationary condition
f  (x) = 0.
However, this stationary condition is just a necessary condition, but it is not a suf-
ficient condition. If f  (x∗ ) = 0 and f  (x∗ ) > 0, it is a local minimum. Conversely, if
f  (x∗ ) = 0 and f  (x∗ ) < 0, then it is a local maximum. However, if f  (x∗ ) = 0 and
f  (x∗ ) = 0, care should be taken because f  (x) may be indefinite (both positive and
negative) when x → x∗ , then x∗ corresponds to a saddle point.
For example, for f (x) = x 3 , we have

f  (x) = 3x 2 , f  (x) = 6x. (1.30)

The stationary condition f  (x) = 3x 2 = 0 gives x∗ = 0. However, we also have

f  (x∗ ) = f  (0) = 0.

In fact, f (x) = x 3 has a saddle point x∗ = 0 because f  (0) = 0 but f  changes sign
from f  (0+) > 0 to f  (0−) < 0 as x moves from positive to negative.

Example 4
For example, to find the maximum or minimum of a univariate function

f (x) = 3x 4 − 4x 3 − 12x 2 + 9, −∞ < x < ∞,

we first have to find its stationary points x∗ when the first derivative f  (x) is zero, that is,

f  (x) = 12x 3 − 12x 2 − 24x = 12(x 3 − x 2 − 2x) = 0.

Since f  (x) = 12(x 3 − x 2 − 2x) = 12x(x + 1)(x − 2) = 0, we have

x∗ = −1, x∗ = 2, x∗ = 0.

The second derivative of f (x) is simply

f  (x) = 36x 2 − 24x − 24.

From the basic calculus we know that the maximum requires f  (x∗ ) ≤ 0 whereas the minimum
requires f  (x∗ ) ≥ 0.
At x∗ = −1, we have

f  (−1) = 36(−1)2 − 24(−1) − 24 = 36 > 0,

so this point corresponds to a local minimum

f (−1) = 3(−1)4 − 4(−1)3 − 12(−1)2 + 9 = 4.

12 Introduction to Algorithms for Data Mining and Machine Learning

Similarly, at x∗ = 2, f  (x∗ ) = 72 > 0, and thus we have another local minimum

f (x∗ ) = −23.

However, at x∗ = 0, we have f  (0) = −24 < 0, which corresponds to a local maximum

f (0) = 9. However, this maximum is not a global maximum because the global maxima for f (x)
occur at x = ±∞.
The global minimum occurs at x∗ = 2 with f (2) = −23.

The maximization of a function f (x) can be converted into the minimization of A−

f (x), where A is usually a large positive number (though A = 0 will do). For example,
we know the maximum of f (x) = e−x , x ∈ (−∞, ∞), is 1 at x∗ = 0. This problem
2

can be converted to the minimization of −f (x). For this reason, the optimization
problems can be expressed as either minimization or maximization depending on the
context and convenience of formulations.
In fact, in the optimization literature, some books formulate all the optimization
problems in terms of maximization, whereas others write these problems in terms of
minimization, though they are in essence dealing with the same problems.

1.3.2 Multivariate functions

We can extend the optimization procedure for univariate functions to multivariate
functions using partial derivatives and relevant conditions. Let us start with the ex-
ample

minimize f (x, y) = x 2 + y 2 , x, y ∈ R. (1.31)

It is obvious that x = 0 and y = 0 is a minimum solution because f (0, 0) = 0. The

question is how to solve this problem formally. We can extend the stationary condition
to partial derivatives, and we have ∂f ∂f
∂x = 0 and ∂y = 0. In this case, we have

∂f ∂f
= 2x + 0 = 0, = 0 + 2y = 0. (1.32)
∂x ∂y

The solution is obviously x∗ = 0 and y∗ = 0.

Now how do we know that it corresponds to a maximum or minimum? If we try to
use the second derivatives, we have four different partial derivatives such as fxx and
fyy , and which one should we use? In fact, we need to define the Hessian matrix from
these second partial derivatives, and we have
⎛ ⎞
 ∂ 2f ∂ 2f
fxx fxy
=⎝ ⎠.
∂x 2 ∂x∂y
H= (1.33)
fyx fyy ∂ 2f ∂ 2f
∂y∂x ∂y 2
Introduction to optimization 13

Since

∂ 2f ∂ 2f
= , (1.34)
∂x∂y ∂y∂x

we can conclude that the Hessian matrix is always symmetric. In the case of f (x, y) =
x 2 + y 2 , it is easy to check that the Hessian matrix is

2 0
H= . (1.35)
0 2

Mathematically speaking, if H is positive definite, then the stationary point (x∗ , y∗ )

corresponds to a local minimum. Similarly, if H is negative definite, then the sta-
tionary point corresponds to a maximum. The definiteness of a symmetric matrix is
controlled by its eigenvalues. For this simple diagonal matrix H , its eigenvalues are
its two diagonal entries 2 and 2. As both eigenvalues are positive, this matrix is pos-
itive definite. Since the Hessian matrix here does not involve any x or y, it is always
positive definite in the whole search domain (x, y) ∈ R2 , so we can conclude that the
solution at point (0, 0) is the global minimum.
Obviously, this is a particular case. In general, the Hessian matrix depends on the
independent variables, but the definiteness test conditions still apply. That is, positive
definiteness of a stationary point means a local minimum. Alternatively, for bivariate
functions, we can define the determinant of the Hessian matrix in Eq. (1.33) as

 = det(H ) = fxx fyy − (fxy )2 . (1.36)

At the stationary point (x∗ , y∗ ), if  > 0 and fxx > 0, then (x∗ , y∗ ) is a local mini-
mum. If  > 0 but fxx < 0, then it is a local maximum. If  = 0, then it is inconclu-
sive, and we have to use other information such as higher-order derivatives. However,
if  < 0, then it is a saddle point. A saddle point is a special point where a local
minimum occurs along one direction, whereas the maximum occurs along another
(orthogonal) direction.

Example 5
To minimize f (x, y) = (x − 1)2 + x 2 y 2 , we have

∂f ∂f
= 2(x − 1) + 2xy 2 = 0, = 0 + 2x 2 y = 0. (1.37)
∂x ∂y

The second condition gives y = 0 or x = 0. Substituting y = 0 into the first condition, we have
x = 1. However, x = 0 does not satisfy the first condition. Therefore, we have a solution x∗ = 1
and y∗ = 0.
For our example with f = (x − 1)2 + x 2 y 2 , we have

∂ 2f 2 2 2
2 + 2, ∂ f = 4xy, ∂ f = 4xy, ∂ f = 2x 2 ,
= 2y (1.38)
∂x 2 ∂x∂y ∂y∂x ∂y 2
14 Introduction to Algorithms for Data Mining and Machine Learning

and thus we have

 
2y 2 + 2 4xy
H= . (1.39)
4xy 2x 2

At the stationary point (x∗ , y∗ ) = (1, 0), the Hessian matrix becomes
 
2 0
H= ,
0 2

which is positive definite because its double eigenvalues 2 are positive. Alternatively, we have
 = 4 > 0 and fxx = 2 > 0. Therefore, (1, 0) is a local minimum.

In fact, for a multivariate function f (x1 , x2 , . . . , xn ) in an n-dimensional space, the

stationary condition can be extended to
∂f ∂f ∂f T
G = ∇f = ( , ,..., ) = 0, (1.40)
∂x1 ∂x2 ∂xn
where G is called the gradient vector. The second derivative test becomes the definite-
ness of the Hessian matrix
⎛ 2 ⎞
∂f ∂ 2f ∂ 2f
...
⎜ ∂x1 2 ∂x1 ∂x2 ∂x1 ∂xn ⎟
⎜ ∂ 2f 2f 2f ⎟
⎜ ∂x ∂x ∂
... ∂x∂2 ∂x ⎟
H =⎜ ⎜ 2 1 ∂x 2 n ⎟. (1.41)
⎟
2
⎜ .. . .
.. . .. .
.. ⎟
⎝ ⎠
∂ 2f ∂ 2f ∂ 2f
∂xn ∂x1 ∂xn ∂x2 ... ∂xn 2

At the stationary point defined by G = ∇f = 0, the positive definiteness of H gives a

local minimum, whereas the negative definiteness corresponds to a local maximum. In
essence, the eigenvalues of the Hessian matrix H determine the local behavior of the
function. As we mentioned before, if H is positive semidefinite, then it corresponds
to a local minimum.

1.4 Nonlinear constrained optimization

As most real-world problems are nonlinear, nonlinear mathematical programming
forms an important part of mathematical optimization methods. A broad class of non-
linear programming problems is about the minimization or maximization of f (x) sub-
ject to no constraints, and another important class is the minimization of a quadratic
objective function subject to nonlinear constraints. There are many other nonlinear
programming problems as well.
Nonlinear programming problems are often classified according to the convexity
of the defining functions. An interesting property of a convex function f is that the
Introduction to optimization 15

vanishing of the gradient ∇f (x ∗ ) = 0 guarantees that the point x∗ is a global minimum

or maximum of f . We will introduce the concept of convexity in the next chapter. If
a function is not convex or concave, then it is much more difficult to find its global
minima or maxima.

1.4.1 Penalty method

For the simple function optimization with equality and inequality constraints, a com-
mon method is the penalty method. For the optimization problem

minimize f (x), x = (x1 , . . . , xn )T ∈ Rn ,

subject to φi (x) = 0, (i = 1, . . . , M), ψj (x) ≤ 0, (j = 1, . . . , N ), (1.42)

the idea is to define a penalty function so that the constrained problem is transformed
into an unconstrained problem. Now we define


M 
N
(x, μi , νj ) = f (x) + μi φi2 (x) + νj max{0, ψj (x)}2 , (1.43)
i=1 j =1

where μi 1 and νj ≥ 0.
For example, let us solve the following minimization problem:

minimize f (x) = 40(x − 1)2 , x ∈ R, subject to g(x) = x − a ≥ 0, (1.44)

where a is a given value. Obviously, without this constraint, the minimum value occurs
at x = 1 with fmin = 0. If a < 1, then the constraint will not affect the result. However,
if a > 1, then the minimum should occur at the boundary x = a (which can be obtained
by inspecting or visualizing the objective function and the constraint). Now we can
define a penalty function (x) using a penalty parameter μ 1. We have

(x, μ) = f (x) + μ[g(x)]2 = 40(x − 1)2 + μ(x − a)2 , (1.45)

which converts the original constrained optimization problem into an unconstrained

problem. From the stationarity condition  (x) = 0 we have
40 − μa
80(x − 1) − 2μ(x − a) = 0, or x∗ = . (1.46)
40 − μ
For a particular case a = 1, we have x∗ = 1, and the result does not depend on μ.
However, in the case of a > 1 (say, a = 5), the result will depend on μ. When a = 5
and μ = 100, we have x∗ = 40 − 100 × 5/40 − 100 = 7.6667. If μ = 1000, then this
gives 50 − 1000 ∗ 5/40 − 1000 = 5.1667. Both values are far from the exact solution
xtrue = a = 5. If we use μ = 104 , then we have x∗ ≈ 5.0167. Similarly, for μ = 105 ,
we have x∗ ≈ 5.00167. This clearly demonstrates that the solution in general depends
on μ. However, it is very difficult to use extremely large values without causing extra
computational difficulties.
Another Random Scribd Document
with Unrelated Content
end, remained practically invisible, was no nearer the fulfilment of its
purpose—the wearing down or rounding up of von Lettow-Vorbeck’s
forces—than it had been when, more than three months earlier, it
had first landed at Port Amelia.
 CHAPTER XVII
THE ADVANCES FROM MEDO TO KORONJE AND
MSALU

All that remained of the Portuguese boma at Medo was the deep
ditch by which it had been surrounded, and the mound or earthwork
fashioned from the earth that had been excavated from it. Any
buildings that these fortifications may have been designed to protect
had long ago been burned to the ground, and save for a big red-
brick store, with an iron roof, situated outside the ditch, there was
no habitable place in the immediate vicinity. It can never have been
of much military value, except against attacks delivered by natives
armed with primitive weapons, and its capture and occupation by
the British conferred upon the latter no material advantage. Medo,
however, or rather the place a few miles east of it where Rock Camp
had been formed, marks the beginning of a stretch of very blind and
difficult country, where big clumps of bamboos are numerous, where
bamboo-brakes of considerable extent are not infrequently
encountered, and where elephant grass nine feet high is a common
feature. Further on along the road, as the columns advanced, more
broken ground was met with, and numbers of isolated rocky hills,
often fantastically shaped—the solitary curved horn of the rhinoceros
being one of the forms most commonly represented—provided the
enemy with excellent observation-posts from which every movement
of the British troops could be watched and provided against.
On the 13th April the two columns camped at Medo, and on the
following day a strong officer’s patrol of the 4th Battalion of the 4th
King’s African Rifles went down the road toward Mwalia, and
speedily found itself engaged with the enemy. Von Lettow-Vorbeck
and Kohl had allowed the British, very slowly and painfully, to work
their way inland from the coast from a distance of eighty-four miles
to Medo; and having now drawn them on into a very difficult belt of
country, they were preparing to ambush the advance once or twice
daily, to make the troops fight as often as possible and in
disadvantageous circumstances, for the camping-ground and for
their supply of water, and to withhold from them any chance of
dealing a very effective blow at their ubiquitous and elusive enemy.
The campaign was at once more harassing and less hopeful than
had been the advance from Narungombe to Lukuledi in the
preceding year, for then “Linforce” had been working its way inland
from Lindi, and there had always been a chance of the enemy being
enveloped by the converging columns; and the country, though thick
and difficult, had not been so blind and so impenetrable as that
through which “Pamforce” was at present engaged in making its
way. Now, too, there was no British force closely co-operating with
“Rosecol” and “Kartucol” to threaten the enemy’s flank and rear,
though some of General Northey’s troops had made their way in a
south-easterly direction from Mahenge, and were known to have
crossed the Rovuma, and Colonel Rose, while still in command in
Portuguese East Africa, had succeeded in getting the 3rd Battalion of
the 2nd King’s African Rifles dispatched to Mozambique, where,
under Colonel Phillips, they were brigaded with a Portuguese force
under Major Leal. There was, however, no immediate prospect of
bringing von Lettow-Vorbeck to a definite action, for there no longer
existed German posts, such as Ruponda, Massassi and Newala, the
defence of which was important to him because their capture would
work him a measure of moral and even of material injury. Instead
von Lettow-Vorbeck, at this time, seemed to have the whole of the
vast continent of Africa into which to retreat, and the prospect of
surrounding or cutting off any large body of his forces was felt by all
to be more remote than ever.
None the less, “Pamforce” continued to move forward down the
road from Medo to Mwalia and from Mwalia to Koronje, with ever-
lengthening lines of communication stringing out behind it, and with
daily ambushes delaying its progress. These, often enough, were
laid for it by small enemy posts consisting of one native non-
commissioned officer and half a dozen Askari, but in such blind
country it was on each occasion necessary to clear up the situation
before the advance could be continued, lest the column should find
themselves caught in some more elaborate trap with results that
might well prove to be disastrous. Moreover, the character of the
country, which greatly favoured the tactics that the enemy was now
adopting, practically confined the British to a series of frontal
attacks, as it did not admit of flanking movements being successfully
carried out.

On the 15th April “Rosecol” left the camp at Medo, and began to
advance down the road in the direction of Mwalia. The 4th Battalion
of the 4th King’s African rifles formed the advanced guard, the Gold
Coast Regiment being in reserve. The former’s advanced points were
attacked, as usual, and the Battalion engaged a small enemy rear-
guard, the progress made during the day amounting to only four and
a half miles. From this time onward, the Gold Coast Regiment and
the 4th Battalion of the 4th King’s African Rifles took it in turns to
lead the advance, and each was preceded at a short distance by an
advanced detachment consisting of 300 rifles with the usual
complement of machine and Lewis guns, and two Stokes guns of the
Gold Coast Regiment. This leading detachment had points thrown
out ahead on each side of the road and a line of skirmishers
deployed behind them, the remainder of the detachment advancing
in open order on both sides of the road, with connecting files
between them and the main body in their rear.
On the 16th April the advanced detachment was supplied by A
Company and two sections of I Company, under the command of
Major Shaw. During the day small engagements were fought with an
enemy rear-guard, consisting of one company, but the Stokes guns
proved very useful and effective, the enemy being shelled out of
successive positions from which, but for these guns, it would have
cost much delay and probably many casualties to eject him. As it
was, only two men of the Regiment and one Sierra Leone carrier
were wounded. The column camped at 2 p.m., Major Shaw’s
detachment digging itself in about a mile further down the road.
On the 17th April the 4th Battalion of the 4th King’s African Rifles
furnished the advanced detachment, that regiment being at the
head of the column, with the Gold Coast Regiment following in
reserve. During the afternoon the King’s African Rifles became
heavily engaged with the enemy, who had been reinforced and was
now opposing the advance with three companies and six machine-
guns. The road here ran through elephant grass nine feet in height,
and it was found impossible to locate the enemy’s positions. On the
other hand, the King’s African Rifles had dug themselves in across
the road, the lie of which was accurately known to the Germans, and
the former consequently sustained many casualties. The 4th
Battalion of the 4th King’s African Rifles was a newly raised force,
largely composed of recruits, and the ordeal of being fired upon by
an invisible enemy, against whom no effective retaliation was
possible, was very severe. However, they held on, and in the
afternoon A Company was sent forward to reinforce them. This
company and the two Stokes gun-teams, which had been with the
4th Battalion of the 4th King’s African Rifles all day, sustained 28
casualties before dark, losing 3 men and 1 battery gun-carrier killed,
and 13 men, 6 battery gun-carriers and 5 Sierra Leone carriers
wounded.
Next morning the Gold Coast Regiment took over from the King’s
African Rifles the position which the latter had occupied during the
night, and was directed to hold the enemy in front while a strong
detachment from “Kartucol” attempted a wide flanking movement on
the right. Captain Duck with thirty rifles was sent forward from the
position held by the Regiment to get in touch with the enemy in
order to give the flanking detachment an objective. He speedily
found and engaged the enemy, whereupon the rest of “Kartucol”
advanced through the Gold Coast Regiment and joined in the fight.
The enemy, however, had once again reduced his rear-guard to a
single company, and on the 19th April “Kartucol” continued the
advance, “Rosecol” following in the rear. On the following day the
two columns were to have exchanged places, but the rations
expected from the rear arrived so late on the night of the 19th April
that this arrangement could not be carried out. The delay had been
caused by the convoy being attacked by the enemy near Rock Camp.
The officer commanding this convoy was killed, and much confusion
was wrought by the ambush, though the carriers and their escort
contrived to get through with the loss of a few bags of mails. There
were many Europeans in camp who would far more willingly have
foregone their dinners. In a captured diary Kohl was subsequently
found complaining with disgust that the mails taken on this occasion
contained no information concerning the progress of the war in
Europe, and mainly consisted of “love to dear Jack.”
"Kartucol," therefore, continued the advance and occupied Mwalia,
while “Rosecol” camped for the night at Kalima, about four miles
short of that place. The distance from Medo to Mwalia is not quite
five-and-twenty miles. The column had left Medo on the 15th April
and “Kartucol” had reached Mwalia on the 20th April, the average
daily progress being therefore little more than four miles.
On the 21st April “Rosecol” remained in camp at Kalima, where it
was joined by General Edwards and his staff. “Kartucol” during the
day was shelled by the enemy, and on the 22nd April it moved
forward and occupied an enemy position two miles in front of the
camp at Mwalia. Both columns remained in these positions until the
26th April, when “Rosecol” moved forward and occupied Makuku,
about twelve miles down the road, “Kartucol,” which had preceded
it, having advanced three miles further to a place called Mbalama. At
Makuku the main road, hitherto followed, which leads from Mtuge to
Lusinje, is crossed by another which runs south-west to Koronje; and
Mbalama is situated some three miles down this latter track.
On the 27th April “Rosecol” advanced through “Kartucol,” and
marched down the road towards Koronje, with Nanungu, some forty
miles further to the west and slightly south of the former place, as
its ultimate objective. The advanced detachment, under Major Shaw,
consisted of the Pioneer Company and A Company of the Gold Coast
Regiment with two Stokes guns. A small party of the enemy was
engaged and driven back; “Rosecol” camped for the night about four
miles west of Mbalama.
 Next day, 28th April, the advance was continued, being led this
time by the 4th Battalion of the 4th King’s African Rifles, two Stokes
guns of the Gold Coast Regiment, as usual, accompanying the
advanced detachment. About six miles were covered during the day,
and as “Rosecol” was forming camp at about 3.30 p.m., patrols from
the advanced detachment came into touch with the enemy, and
Lieutenant McEvoy was wounded in the hand by a stray bullet, and a
trumpeter belonging to the Stokes Gun Battery was killed.
On the 29th April the enemy was found to have abandoned the
positions which he had occupied the night before; and at 7 a.m. the
advanced detachment, consisting of half I and B Companies with
two of the Gold Coast Stokes guns, advanced, the rest of “Rosecol”
following half an hour later. Major Shaw, who was, as usual, in
command of the advanced detachment, came into contact with the
enemy at about 10.30 a.m., and thereafter the latter fought an
intermittent rear-guard action—a series of harassing ambushes—
until 4.30 p.m., when camps were formed for the night, Major
Shaw’s men occupying a position about a mile in advance of the rest
of the column. In the course of the day only two men of the Gold
Coast Regiment were wounded, the Stokes guns once more proving
very useful in dislodging the enemy from successive positions.
On the 30th April, “Kartucol” passed through “Rosecol” with the
intention of attacking an enemy position, which was known to be
held by four companies and one gun. The Headquarters of the Gold
Coast Regiment, with half the Stokes Battery, the Pioneers and I
Company, marched in the rear of “Kartucol” as reserve troops. Touch
was not gained with the enemy until the afternoon, but owing to the
country traversed being very difficult and blind, the progress made
was so slow that no attack could be delivered upon the German
position owing to the lateness of the hour. The two columns,
therefore, formed a perimeter camp at about 4.30 p.m. at a place on
the Koronje road about four hundred yards west of the Montepuez
River. One Battalion from “Kartucol” occupied an advanced camp
about one thousand yards further down the road leading to Koronje.
On the 1st May, the 1st Battalion of the 2nd King’s African Rifles
advanced along the road toward Koronje, while the 2nd Battalion of
the same Regiment went out on the right to attempt to outflank the
enemy’s left. The country was still very difficult and extremely blind,
and progress was again very slow. It was subsequently discovered,
moreover, that from an observation post on the summit of Koronje
Hill, to the left of the road, the enemy could follow every movement
of the British troops. While, therefore, the 2nd Battalion of the 2nd
King’s African Rifles was laboriously working its way round to the
right, its attempt to surprise and outflank the enemy was
foredoomed to failure from the outset. Meanwhile, of course, this
movement greatly delayed the advance of the rest of the force.
The detachment of the Gold Coast Regiment which, under the
command of Major Shaw, was with “Kartucol,” was employed to
escort the 22nd D.M.B. and the ammunition column of that force.
At about 5 p.m. the 1st Battalion of the 2nd King’s African Rifles
became heavily engaged, and simultaneously an enemy party of
about forty rifles, which had worked its way through the bush to the
rear, attacked the D.M.B. which was being escorted by fifty rifles of I
Company. The latter, under Lieutenant Kay, acted with great
steadiness and promptitude. At the moment when the attack was
delivered, the Mountain Battery, which had just come out of action,
was limbered up. For a moment the guns were in peril, but
Lieutenant Kay held the enemy and beat off the attack while the
mules and their loads were got away in safety.
The sound of the firing misled the 2nd Battalion of the 2nd King’s
African Rifles, which was out on the right, with the result that it
rejoined the column in the rear of the enemy.
A perimeter camp was formed for the night, the 1st Battalion of
the 2nd King’s African Rifles digging themselves in at a point about
eight hundred yards in advance of the main body.
On the 2nd May, the 1st Battalion of the 2nd King’s African Rifles
pushed out patrols which quickly came into touch with the enemy,
who was soon after engaged by “Kartucol,” which drove him back.
No progress, however, was made during the day, and on the morrow
it was found that, while the enemy’s rear-guard was fighting
“Kartucol,” the position at Koronje had been evacuated. “Kartucol”
then advanced and camped near Koronje, the detachment of the
Gold Coast Regiment under Major Shaw rejoining “Rosecol” in the
afternoon.
On the 4th May “Kartucol” again advanced and located a strong
enemy position near the Milinch hills, about six and a half miles west
of Koronje, through which the road passes. On this day three officers
and ten British non-commissioned officers belonging to the Gold
Coast Regiment arrived from Port Amelia.
On the 5th May, “Rosecol” advanced and took over from
“Kartucol,” which then fell back to the camp which the former had
hitherto occupied. The 4th Battalion of the 4th King’s African Rifles
encamped at a point down the road about a mile in advance of the
main body of “Rosecol”; and patrols were sent out to the right and
left to try to find a way round the enemy’s position on the Milinch
Hills. Both these patrols were furnished by A Company of the Gold
Coast Regiment, that on the right being commanded by Captain
Harris and that on the left by Lieutenant Withers.
On the 6th May Captain Harris returned and reported that the
country to the north was much more open than that through which
the columns had recently been advancing, and that it would be
almost impossible to make a flanking movement from the right side
of the road. On the 7th May Lieutenant Withers came in from the
south bringing a similar report; and meanwhile patrols sent out by
the 4th Battalion of the 4th King’s African Rifles had on both days
come into touch with the enemy just east of the Milinch Hills, and
reported that the position which he was occupying was a very strong
one. This was indeed the case, for the enemy was posted on the
crests and slopes of two hills, both of which commanded the gut
between them through which the road runs; yet on the 8th May it
was discovered that the Germans had retired, and two companies of
the 4th Battalion of the 4th King’s African Rifles went forward and
occupied the position which he had evacuated.
Meanwhile the lines of communication were lengthening behind
the columns, and now measured approximately one hundred and
forty miles from Mtuge, which in its turn is twenty-eight miles by
road from Port Amelia. Also the heavy and increasing traffic over the
road had not tended to improve it; and though road corps, recruited
from South Africa and East Africa, toiled ceaselessly at its repair, the
difficulties of transport and supply were becoming daily more and
more acute. At this time, the columns at the front had been on very
short commons for a considerable period, and the company officers
of the Gold Coast Regiment reported that their men were not getting
enough food to keep them fit to take part in active operations of so
trying and arduous a character as those at present in progress.
On the 9th May the Gold Coast Regiment took over the Milinch
Hills from the 4th Battalion of the 4th King’s African Rifles; and on
this day local natives reported to Colonel Rose that von Lettow-
Vorbeck, with a large enemy force, was moving in a north-easterly
direction toward Lusinje. This place lies about thirty-seven miles
almost due north of Nanungu, on the main road from which the
columns had branched off in a south-westerly direction at Makuku,
as already noted. Accordingly the 4th Battalion of the 4th King’s
African Rifles was dispatched across country to Msalu Boma, which is
situated on that road at a point, as the crow flies, about twenty-
three miles north-west of Koronje, and twenty-seven miles east by
south of Lusinje. The orders issued to this battalion of the King’s
African Rifles were that they should deal with any enemy parties
weak enough to enable action to be taken with effect, but to avoid
any serious engagement with his numerically superior forces.
It was believed that a fairly strong party of the enemy were
occupying a hill on the right side of the road at a place called
Jirimita, about five or six miles down the road from the pass through
the Milinch Hills, and at dawn on this day two patrols were sent out,
one under the command of Captain Leslie-Smith and the other under
Lieutenant Bisshopp. Each patrol consisted of seventy-five rifles,
drawn respectively from A and I Companies; and Captain Leslie-
Smith, who went out on the right of the road, had orders to make a
flanking movement and to come back to the highway at a point
about four miles beyond Jirimita. Lieutenant Bisshopp, on the left,
was instructed to make a wider and longer sweep, and to strike the
road about three miles further on. It was hoped thus to outflank the
enemy and to cut off his retreat. It was a difficult task in the broken
country through which these two patrols had to work, at once to
maintain a correct sense of direction, and accurately to estimate the
distance traversed. However, both these small parties started off,
expecting to be a night or two in the bush, and each in the end
succeeded in exactly carrying out the orders issued to it.
Meanwhile, during the morning of the same day Lieutenant
Wilson, with a patrol of twenty rifles drawn from the Pioneer
Company, got touch with an enemy outpost of about the same
strength at a point some two miles west of the Milinch Hills; and at
4.45 p.m. a second officer’s patrol, under Lieutenant Beech, was
sent out down the road in the same direction for a distance of two
and a half miles without coming into contact with the enemy, whose
outpost had retired since the morning.
At 6 a.m. on the 10th May, Lieutenant Withers, with fifty rifles and
one Lewis gun of A Company, was sent down the road with orders to
brush aside any small party of the enemy that he might encounter,
and thereafter to try and ascertain the real strength of the force
which was opposing the advance of the column.
Three and a half miles from the Milinch Hill Lieutenant Withers
met a small party of the enemy, which he drove back; and about five
miles out he found an enemy camp, strongly entrenched, which had
evidently been designed to accommodate some four companies, but
which had been recently burned. As far as it was possible to judge,
this camp had been destroyed and abandoned two days earlier; and
though the tracks leading from it were at once confused and
confusing, conveying at first the impression that the enemy had
retired in a northerly direction, it was subsequently ascertained that
he had retreated down the main road. Just beyond the burned camp
this road was found to bifurcate, one fork leading west-north-west
and the other west-south-west. It was the latter route which the
enemy had taken.
The main patrol camped at a point where the road bifurcated, and
sent out small parties to reconnoitre along each of the forks, but
neither of them came into touch with the enemy.
On the 11th May the patrols under Captain Leslie-Smith and
Lieutenant Bisshopp, which had been sent out on the 9th May,
rejoined the Regiment. As has already been noted, they had
achieved the difficult feat of striking the road at the points aimed at,
but for the rest, though Lieutenant Bisshopp’s patrol had surprised
and killed one enemy Askari, who had probably been left behind to
watch the movements of the British, nothing more had been seen of
the enemy, who must have passed down the road while these
patrols were still making their way through the bush.
On the 12th May one of the battalions of the 2nd King’s African
Rifles from “Kartucol” took over from the Gold Coast Regiment,
which returned to the main camp occupied by “Rosecol.” On the
following day the latter marched across country, in the wake of the
4th Battalion of the 4th King’s African Rifles, which had preceded
them on the 9th May, in the direction of Msalu Boma. The way led
along a native footpath which only admitted of men marching in
single file, but in order to beat out a track for the transport through
the high grass and standing crops of maize and millet, the column
advanced four abreast—a hard task for troops who had been
insufficiently fed for many days, and who were now required to
cover between daybreak and dusk a distance of eighteen miles. The
column camped in the bush, and on the following day it joined up
with the 4th Battalion of the 4th King’s African Rifles at the boma at
Msalu. This place, too, had once been a stronghold of a Portuguese
revenue-farmer, and had been fortified against attack by the natives,
but it had now been completely destroyed by fire.
At Msalu news was received that von Lettow-Vorbeck and the
whole of his main force were at Nanungu, and that so far they had
given no signs of any intention to move to the north toward the
Rovuma River, or south to the Lurio, which divides the territory of
the Nyassa Company from the Province of Mozambique. It was also
learned that the King’s African Rifles Mounted Infantry were at
Lusinje, some six-and-twenty miles along the main road west by
north of Msalu, and about thirty-two miles almost due north of
Nanungu.
“Rosecol” remained at Msalu on the 14th and 15th May, the
neighbourhood being clear of the enemy, but much infested by lions.
The proximity of these brutes got upon the nerves of some of the
inmates of the camp, and on the night of the 13th—14th May a
carrier, who had had a nightmare in which they played a prominent
part, awoke in a panic, shattering the silence with his yells and
outcry. Instantly an indescribable scene resulted. Tumbling over one
another to get at the camp-fires, the porters fought and scrambled
for firebrands which they waved wildly, and impeded by which they
made desperate efforts to climb into neighbouring trees. The country
here is orchard-bush, and the only trees available are small and
stunted—altogether inadequate as places of refuge from the
onslaught of a lion. The terrified carriers, however, were long past
reason, and appeared to consider that their one chance of salvation
lay in getting even a foot or two above the ground. The lions on this
occasion existed only in their imagination, and order and confidence
were presently restored. During the same night, however, the 4th
King’s African Rifles lost two sentries, one killed and one badly
mauled by these brutes, so the terror of the carriers had at any rate
some measure of justification.
With the arrival of “Rosecol” at Msalu the second phase of the
advance, which had its beginning with the fight at Medo, may be
regarded as concluded. The enemy had offered a persistent and
fairly effective resistance to the progress of the columns along the
main road through the difficult country which lies between Medo and
the Milinch Hills. His main force, which was believed to be at
Nanungu, was really encamped at Wanakoti, about three and a half
miles to the north of that place; and against him were advancing
“Kartucol” from the east, “Rosecol” from the north-east, and a weak
column of perhaps 800 rifles, which General Northey had dispatched
across the Rovuma in a south-easterly direction, under the command
of Colonel Griffiths. Von Lettow-Vorbeck still had the choice of
several lines of retreat, for at Wanakoti many tracks cross one
another, and though the road to Koronje on the east and to Chisona
on the north-west were closed to him by the British advance, the
track leading south-west to Mahu was still open, and while retreating
along it he would have opportunities of breaking off, should it suit
his convenience to do so, in almost any direction.
 CHAPTER XVIII
THE EXPULSION OF VON LETTOW-VORBECK FROM
THE NYASSA COMPANY’S TERRITORY AND THE
RETURN OF THE GOLD COAST REGIMENT

On the 16th May “Rosecol” left Msalu, and marching along bush paths
in a westerly direction, leaving the road to Lusinje on the north and
having the Msalu River on its right, began a movement which was
designed to cut the main road between Lusinje and Nanungu. Camping
for two nights in the bush—orchard country which, though the soil was
of a rocky character, was broken by frequent patches of cultivated land—
the column crossed this road on the 18th May, and pushed on toward
Chisona. On reaching the Lusinje–Nanungu road, a patrol was dispatched
to examine the ford across the Msalu River, and on approaching it was
fired upon by a party which proved to be composed of scouts belonging
to the Rhodesian Native Regiment—part of the weak column which
General Northey had sent out across the Rovuma River. Connection was
thus established for the first time with this force.
On the 19th May “Rosecol” continued its march to Chisona, where it
camped on the banks of the Msalu River at a place about two miles from
the column from “Norforce” above mentioned, which was under the
command of Colonel Griffiths. The river was unfordable at this season of
the year, but the battery-carriers quickly constructed a bridge under the
personal supervision of Colonel Goodwin, who, as a former commander
of the Pioneer Company, had proved himself, both in the Kameruns and
in East Africa, to possess a special gift for such improvizations.
On the 20th May the column crossed the Msalu, and marched due
south to within five miles of Chilonga, I Company leading the advance
and doing what it could to widen and improve the existing paths so as to
facilitate the passage of the column. On the 21st May the latter pushed
on twelve miles in a westerly direction and camped at a spot some three
miles to the north of the road to Mahua. Five companies of the enemy,
under Kohl, were reported to be on this road; and it was here learned
that “Kartucol” had entered and occupied Nanungu without opposition,
and was advancing along the Mahua road. This advance had been
opposed by Kohl during the day, one company of the enemy with one
gun having been in action, while the rest of his force was held in reserve.
Meanwhile Colonel Griffiths’ column was marching parallel to “Rosecol,”
on a line a few miles to the north of it.
At this juncture General Edwards hoped to surround Kohl from the
west, east, and north; and with this object in view “Kartucol” was
ordered to advance along the Mahua road, Colonel Griffiths’ column to
march in a south-westerly direction, so as to get astride that road in the
rear of the enemy, while “Rosecol” was instructed to march on a line
about three miles to the north of the Mahua road and roughly parallel to
it with the object of turning the enemy’s left.
During the afternoon of the 22nd May Colonel Griffiths’ force was
heard to be heavily engaged, and “Rosecol” continued its march until 10
p.m., when it camped, Major Shaw in command of the Pioneers and B
Company of the Gold Coast Regiment, and two Gold Coast Stokes guns,
forming an advanced detachment encamped on high ground a few miles
forward, overlooking the place where Colonel Griffiths was entrenched.
During all these operations “Rosecol” was separated from “Kartucol” by
the Mwambia Ridge—a high barrier of grey, granite hills, with unscalable,
cliff-like sides, rising abruptly from the grass and bush and orchard forest
at their base—which flanks the main road on the north for a matter of
more than a dozen miles.
Colonel Griffiths’ column, it appeared, had struck the Mahua road, and
had entered and occupied Kohl’s camp at Mwariba, meeting with very
little resistance. Here he had possessed himself of practically all Kohl’s
heavy baggage—a really severe loss to the enemy at this juncture; but
almost immediately afterwards he had been vigorously attacked, his
small column being completely surrounded and suffering many
casualties. Failing to push home his attack, however, the enemy had
drawn off during the night and had then retired in a southerly direction.
Yet another attempt to envelop him had definitely failed.
The Gold Coast Regiment this day came into contact with the enemy
for the first time since it had quitted the main road near Koronje on the
13th April. Its only casualty, however, was one man wounded.
On the 23rd May “Rosecol” advanced through Colonel Griffiths’ camp,
with Major Shaw’s detachment about one mile ahead of it; and very
shortly afterwards the latter became engaged with the enemy, who, with
one company and two machine-guns, was covering the retirement of
Kohl’s main force. Major Shaw drove this enemy party back a matter of
two miles, when he was relieved by the 4th Battalion of the 4th King’s
African Rifles, who now formed the advanced detachment of “Rosecol,”
supported as usual, however, by two guns of the Gold Coast Regiment’s
Stokes Battery.
On this morning the Regiment lost one British non-commissioned
officer, Sergeant Kent, and one soldier killed, and three men wounded.
On the 24th May the 4th King’s African Rifles advanced at 6 a.m., and
forthwith became engaged with the enemy, whose strength had now
been increased to at least two companies with four machine-guns. All
day long the Germans fought a series of very stubborn rear-guard
actions, and the progress made by dusk was only two miles. In the
course of the day Lieutenant Percy and two battery gun-carriers,
attached to the Gold Coast Stokes guns, were wounded.
On the 25th May “Rosecol” advanced along the Mahua road in the
direction of Korewa, with “Kartucol” following in its rear; Colonel Griffiths’
column having marched west on the preceding day with the object of
once again getting astride the road behind the enemy, this time on the
other side of Korewa. The enemy was not met with, however, Major
Shaw occupying Korewa in the afternoon without opposition, and during
the night news was received that Colonel Griffiths had struck the road at
the point aimed at, and that he, too, had seen nothing of the enemy.
From Korewa patrols were sent out in several directions, and by the
27th May, it having by then become pretty evident that von Lettow-
Vorbeck with the main body, followed at a short distance by Major Kohl
and his redoubtable rear-guard, had crossed the Lurio River into the
province of Mozambique, Colonel Griffiths’ column marched that evening
in pursuit.
On the 28th May B Company, less one machine-gun and one Lewis
gun, left the camp at 6 a.m. for Wanakoti, thirty miles to the east, acting
as escort to the 22nd D.M.B. The rest of the Regiment remained in camp
at Korewa, where it was rejoined by B Company in due course.
 With the retreat of von Lettow-Vorbeck southward across the Lurio
River, the expedition into the Nyassa Company’s territory, which had
been begun five months earlier by the landing of Major Shaw’s advanced
detachment at Port Amelia, reached its natural termination. Yet another
campaign, based so far as the British were concerned upon the port of
Mozambique, was about to begin, though as yet no very extensive
preparations had been made for its effectual initiation.
The Gold Coast Regiment, as it has been seen, had been transferred
straight from the pursuit of von Lettow-Vorbeck through the Kilwa and
Lindi areas and on to the banks of the Rovuma, to the very trying inland
march from Port Amelia. Other units subsequently engaged in that
enterprise had in the interval been afforded a period of rest, the 2nd
Battalion of the 2nd Regiment of the King’s African Rifles, for example,
having been allowed to return for a space to their cantonments and to
their womenkind at Nairobi. The men of this corps and those of the Gold
Coast Regiment, who had done so much hard fighting in company, had
learned greatly to trust and value one another, and though they were
drawn from such widely different parts of the African continent and
though the Gold Coast soldiers’ knowledge of Swahili was still rather
elementary, a species of blood-brotherhood had come to be recognized
as existing between them. When the “Second Second,” as this battalion
of the King’s African Rifles was familiarly called, had made its appearance
in Portuguese East Africa, it had been warmly welcomed by the men of
the Gold Coast Regiment, and the latter, it may be surmised, had listened
not without envy to the accounts which their friends had to give them of
the good time the former had enjoyed during their stay at Nairobi. Were
the war-worn veterans of the Gold Coast Regiment never to enjoy a
similar respite from patrols, attacks, counter-attacks and endless toils
and fatigues? The men put the question to their officers. They would
fight on if they must, embarking forthwith upon this new campaign
which was clearly about to begin; but they would fight better, they felt, if
in the interval they might have a taste of the delights of rest and home
in their cantonments at Kumasi. Colonel Goodwin, who was now
commanding the Regiment, and Colonel Rose, who was commanding the
column to which the battalion was attached, shared the men’s opinion,
and General Edwards agreed that the Regiment had fairly earned a rest.
Accordingly, at 7 a.m. on the 1st June, the Gold Coast Regiment left
the camp at Korewa, and began its march back to Port Amelia. From
Medo to Ankuabe—a distance of five-and-twenty miles—it was conveyed
by motor-cars, but the rest of that weary journey was accomplished on
foot over a road which had been knocked to pieces by the traffic passing
over it. A standing camp was established at Gara, between Mtuge and
Bandari, which was reached on the 13th June, Colonel Rose having, on
the preceding day, relinquished the command of “Rosecol” and resumed
that of the Regiment.
The rest of June, July and the first twelve days of August were spent
in refitting, and men of the Regiment who were doing duty at various
points along the lines of communication were gradually recalled and
collected. On the 29th July Colonel Rose and Major Read sailed for South
Africa from Port Amelia on board H.M. Transport Hymettus; and on the
13th August Major Hornby with 37 officers, 17 British non-commissioned
officers, 862 rank and file, and 135 stretcher-bearers, gun-carriers, etc.,
embarked on board H.M.T. Magdalena and on the 14th August set sail
for West Africa.
At Durban, reached on the 18th August, Colonel Rose and Major Read
rejoined the Regiment, and both here and at Capetown, where the
transport arrived on the 27th August, several officers were landed who
were taking leave in South Africa, Australia or Tasmania.
Accra, the capital of the Gold Coast, was reached without incident late
on the 5th September, and on the following day the Governor, who had
seen the Regiment off from Sekondi exactly two years and two months
earlier, came on board the Magdalena to welcome and inspect the
troops, and to thank them on behalf of the Colony whose name they
bear, for the splendid fashion in which, through all the trials and dangers
of the East African campaign, they had upheld its reputation.
Colonel Rose and Major Read disembarked at Accra, but the Regiment
sailed on the evening of the 6th September for Sekondi, where it arrived
early next morning.
From this port to Kumasi, whither the Regiment at once proceeded in
special trains, its journey was a triumphal progress. At Sekondi itself a
feast of native foods, such as these soldier-exiles had not tasted for two
years, had been prepared for their consumption; and at every halting-
place crowds had assembled to greet and acclaim the Regiment and to
load the men with gifts. All along the line little knots of natives shouted
and danced their welcome, and even after darkness had fallen every
station at which the trains stopped was crammed by eager crowds of
Europeans and natives alike, bent upon showing the men what pride the
colony felt at the reputation which they had won for themselves, and
how deep was the popular sympathy for all they had suffered and
endured.
It was a royal home-coming, and when at dawn the men, worn out
with excitement and fatigue at last arrived at Kumasi, their women met
them at the station in a clamorous mob, and accompanied them in
triumph to their cantonments, with the songs and dances wherewith the
warriors of West Africa have always been greeted on their return from a
victorious campaign.
But, alas! there were wailings and keenings too, mingling with the
joyful tumult, for many a woman there was lamenting some poor fellow
who lies buried far away on the other side of Africa, and would not be
comforted because he was not.

The casualties sustained by the Gold Coast Regiment during the

campaign in East Africa were as follows:—
Killed
in Died of
action. Wounded. Missing. diseases. Invalided.
British officers 9 21 —
British non-
commissioned
officers 6 9 — 4 15
Rank and file 181 603 13 206 469
Gun-carriers 9 56 — 16 28
Stretcher-
bearers — — — —
Clerks — — — 1 1
Carriers 10 33 — 40 24
Total 215 725 13 270 567
The strength of the Gold Coast Regiment actually in the field never
much exceeded 900 rifles. The total of effectives belonging to the
Regiment at any one time in East Africa never numbered much more
than 3000, and from first to last the total number of officers and men of
all ranks dispatched did not amount to much more than 3800. When
these facts are remembered, the above table will be found strikingly to
illustrate the severity of the fighting in which the Regiment took so active
a part, and to indicate the ravages caused by disease to which prolonged
strain and hardship exposed it.

Meanwhile the recruiting efforts made by the Government of the Gold

Coast, to which during 1917-18 Captain Armitage, C.M.G., D.S.O., the
Chief Commissioner of the Northern Territories, had devoted special
energy and enthusiasm, had resulted in the collection of a very large
number of recruits at the various training-depôts throughout the Colony,
Ashanti, and the Northern Territories; and the Regiment had proved itself
to possess such fine qualities that, as the early end of the war was not at
that time anticipated, the War Office decided to convert it from a
battalion to a brigade. This consisted of four full battalions with a battery
of four 2·75 guns, and a battery of eight Stokes guns, and it was
constituted a brigade as from the 1st November, 1918, under the
command of Brigadier-General Rose. It was an open secret that, as soon
as its organization was complete, the Second West African Brigade, as it
was now called, was to be dispatched on active service to Palestine.
Then, during the closing days of October and the first half of
November, came the dramatic collapse of the Central Powers and of their
Allies—the débacle in the Balkans, the surrender of Turkey, the rout of
the Austro-Hungarian armies on the Italian front, the succession of
hammer-blows delivered on the western front from the Swiss frontier to
the sea, and finally the Armstice granted to a defeated, crime-stained
enemy, the terms of which exactly reflected the magnitude of the Allies’
victory, and the extent to which Germany and Germans had forfeited the
trust and the respect of all mankind.
The reading of those terms from the balcony of the Public Offices at
Accra to a large concourse of people, almost beside themselves with
enthusiasm and delight, was recognized as closing the short career of
the Gold Coast Service Brigade; and by the end of the following
December its disbandment was completed. It had existed long enough,
however, to enable the Gold Coast to boast that it, no less than its
neighbour the huge territory of Nigeria, had been able to raise by
voluntary enlistment a full brigade of soldiers for the service of the
Empire in the Great War.
 APPENDIX I

THE MOUNTED INFANTRY OF THE GOLD COAST| REGIMENT

There is another Gold Coast unit, which never served with the rest
of the Regiment, and which remained behind in Portuguese East
Africa when the remainder of the battalion returned to the West
Coast, and of its short but adventuresome career some brief account
must here be given.
At the end of February, 1918, nearly two months after the arrival
of Colonel Goodwin with the main body of the Gold Coast Regiment
at Port Amelia, Lieutenant G. H. Parker, who has been mentioned in
an earlier chapter as having been in temporary command of the
Battery, was chosen by Colonel Rose to raise and train a small body
of Mounted Infantry. He was told to pick out for this purpose, from a
newly arrived draft of recruits from the Gold Coast, 170 men; and to
him were attached Lieutenants Drummond and Saunders, and five
British non-commissioned officers.
The men chosen were natives of the Hinterland of the Gold Coast,
to whom, since they for the most part live beyond the range of the
tsetse fly and the Trypanosoma, horses are more or less familiar
animals. About 10 per cent. of them could ride in the hunched-up,
Tod Sloan-like fashion peculiar to folk to whom saddles are unusual
luxuries; but not a man among them had the vaguest ideas
concerning horse-mastership and management.
Four riding-schools were constructed near the camp, upon the top
of the hill which slopes on the one side to the waters of the Indian
Ocean, and on the other falls in a sheer cliff to the beach at Port
Amelia; and daily for hours at a time the European officers and non-
commissioned officers shouted themselves hoarse, while the men
bumped round the manèges. A certain number were incurably
horse-shy, and had to be “returned to store,” but the majority were
quite fearless and enjoyed their daily ride, and though horses had
not been received at Port Amelia until the end of March, by the 30th
May No. 1 Troop of the Mounted Infantry of the Gold Coast
Regiment was declared to be fit to take the field.
This troop, under the command of Lieutenant Drummond,
consisting of 1 British non-commissioned officer, 41 rank and file, 51
horses, 2 mules, and 2 camp-followers, left Port Amelia on the
above-mentioned date, and rode up the well-worn track from Mtuge
to Medo, and thence to Wanakoti, General Edwards’ Headquarters.
The troop arrived at this place just as the Gold Coast Regiment was
about to begin its march back to the coast from Korewa.
It is not possible to follow the history of this troop in detail without
embarking upon a full account of the campaign in the Province of
Mozambique, to which the British were committed after the Germans
had retreated across the Lurio River, and this forms no part of the
plan of the present work. It must, therefore, suffice to note that
“Kartucol” from this time onward followed hard upon the heels of the
enemy forces, pursuing them without intermission nearly as far
south as Kilimane. A little north of this place one and a half
companies of the 2nd Battalion of the 3rd King’s African Rifles, with
a much larger force of Portuguese encamped at Nhamaccura, were
attacked by the enemy, who, having possessed himself of the guns
belonging to the Portuguese, nearly annihilated the small British
detachment, Colonel Gore Brown, who was in command, being
himself killed with a large number of his men.
After this the enemy went north once more, still pursued by
“Kartucol,” which had now cut loose from its transport and was living
on the country; and the Germans shortly afterwards attacked and
invested Namirrue, a place near the centre of the province, which
was being held by a company of the 2nd Battalion of the 3rd King’s
African Rifles, under Captain Bustard.
 Drummond’s Troop of Gold Coast Mounted Infantry had worked its
way down in a southerly direction from Wanakoti to Namirrue,
scouting for the columns, and doing some excellent work; and it had
joined up with Captain Bustard’s little force just before the latter was
surrounded.
Finding that the position which he occupied at the moment of the
enemy attack was commanded by the German guns, Captain
Bustard occupied a higher hill near at hand; and, though hopelessly
outnumbered, cut off from water, and bombarded by a Stokes gun
which had been captured by the Germans, he made a gallant fight of
it, and held out for three days.
Meanwhile the three remaining troops of the Gold Coast Mounted
Infantry, under Captain Parker, had sailed from Port Amelia on the
1st July, arriving on the following day at Mussuril Bay, in the
entrance to which lies the island of Mozambique. The force consisted
of 8 British officers, 10 British non-commissioned officers, 137 rank
and file, 84 East Africans, 2 Indians, 11 other details, with 133
horses, 50 mules, and 141 donkeys.
The Mounted Infantry were disembarked at Lumbo, on the
northern shores of the bay, and on the 5th July marched twenty
miles to Monapo, where their depôt was established. On the 8th July
the Squadron began its march to Nampula, eighty miles further
inland, where at this time General Edwards had his Headquarters;
and travelling an average of about twenty miles a day, it reached its
destination on the afternoon of the 11th July. Here Captain Parker
learned that No. 1 Troop was with Captain Bustard at Namirrue, and
that it was thought that the small post established there would
embarrass the retreat of the enemy, who was known to be
advancing from the south.
On the following days the Squadron pushed on in the direction of
Chinga, which lies five-and-forty miles to the west of Nampula,
walking and leading most of the time, for sore backs among the
horses were already giving occasion for anxiety. From Chinga on the
15th July the Squadron marched sixteen miles to Marrupula; and
here on the following day Captain Parker received orders to press
forward as rapidly as possible to Metil, and thence to take up certain
positions on the Ligonha River. Three days’ rations were drawn, and
though the nights were very cold, the capes and spare clothing were
all left behind, the men being cut down to their body-clothes and
one blanket each, in order to ease the horses of as much weight as
possible.
On the 17th July the Squadron covered a distance of thirty-three
miles to Calipo, and on the morrow reached Pequerra, and pushed
on thence to the banks of the Ligonha River, travelling on that day
thirty-six miles between dawn and dusk.
The geography of the country was very imperfectly known, and
the only available maps were grossly inaccurate. Moreover, whereas
it had been anticipated that the Ligonha River would only be
fordable in a few places, which the Squadron had been ordered to
hold, it was found that the stream was quite shallow for a distance
of at least twenty miles. This was discovered on the 19th July, on
which day Metil was reached, the Squadron having marched one
hundred and two miles to that place from Murrupula in fifty-seven
hours—a very good performance for a newly raised body of Mounted
Infantry.
From Metil one troop, under Lieutenant Poole, was sent eastward
to Napue; a second, under Lieutenant Viney, went toward
Muligudge, five miles south-east of Metil; and a third, under
Lieutenant Saunders, back along the track towards Pequerra, twenty
men and Lieutenant Broomfield remaining at Metil with Captain
Parker. All these mounted patrols had orders to try to locate the
enemy and to keep touch as far as possible with one another and
with Captain Parker.
On the 23rd July news was received that Namirrue was invested
by the enemy, and that though it was still holding out, Colonel
Fitzgerald’s column, consisting of the 4th Battalion of the 4th and the
3rd Battalion of the 3rd King’s African Rifles, had had to retire when
attempting to move to Captain Bustard’s relief. As Captain Parker
was instructed to get as many of his men together as possible in
order to scout in the direction of Namirrue, the troops under
Lieutenants Poole and Viney were recalled, and on the 24th July,
Captain Parker moved back to Pequerra, and thence proceeded
through dense bush to the banks of the Ligonha. From here
Lieutenant Viney with twelve troopers crossed the stream and went
scouting in what was believed to be the direction of Namirrue. On
the 26th July Lieutenant Broomfield with twenty men were sent to
Lulete, Captain Parker and Lieutenant Saunders with twenty-eight
men—all that remained at their disposal—moving up the left bank of
the Ligonha. At 4 p.m. they came upon a track, surprised an enemy
baggage-train, and captured nine porters, the baggage-guard
making off. Lieutenant Saunders with a few rifles was left to watch
the trail, and late that afternoon he had a brush with the enemy, and
captured a German, an Askari, and about a dozen more porters, also
killing one or two enemy soldiers.
On the 27th July Captain Parker set off for Pequerra with the
prisoners, leaving Lieutenant Saunders with a few men to watch the
track and to snipe and harass the enemy. Captain Parker fell in with
a superior force of the enemy, lost all his prisoners and a good many
of his men and horses, and was himself reported missing for three
days. At the end of that time, however, he and the surviving
remnant of his troop contrived to rejoin. Meanwhile Lieutenant
Saunders also came into touch with the enemy, was wounded and
had several of his men and nearly all his horses shot; while
Lieutenant Viney, who was surprised and attacked just as he had off-
saddled, was killed, his men, acting on his orders, dispersing into the
bush. Immediately afterwards word was received that Captain
Bustard at Namirrue had been compelled to surrender, and with him
Lieutenant Drummond and what was left of No. 1 Troop of the Gold
Coast Mounted Infantry.
This meant that the Gold Coast Mounted Infantry, which on the
5th July had numbered, including Lieutenant Drummond’s troop,
about one hundred and sixty-five rank and file, was now reduced to
sixty-five men; and Captain Parker returned to the depôt at Mnapo
to train and equip further drafts, while Lieutenant Broomfield
remained in the field in command of the handful of mounted men
still effective as a fighting force.
Though the enemy had won successes at Nhamaccura and at
Namirrue, in both of which places he had succeeded in cutting off
small British forces, he was now being hunted by “Kartucol” from the
south into the grip of six converging columns; and for the first time
in the history of the whole campaign he was so completely cornered
that in the neighbourhood of Chalana—a place some five-and-forty
miles inland from the coast of Antonio Annes—he was compelled to
concentrate all his troops, combining them into a single force.
It was while the meshes of the net appeared at last to be securely
drawn around von Lettow-Vorbeck, that Lieutenant Broomfield and
his little body of sixty men of the Gold Coast Mounted Infantry
specially distinguished themselves. It was of great moment to
General Edwards that he should be kept fully and frequently
informed of the exact position and movements of the enemy, and
this service was rendered to him by Lieutenant Broomfield. For a
week the Gold Coast Mounted Infantry maintained close touch with
the enemy’s main body. The country is here very thickly populated.
The Germans, who were paying for all their supplies with cloth which
they had looted from the Portuguese stores, had made themselves
very popular with the local natives, who witnessed the wholesale
destruction of the Portuguese bomas with ecstatic delight. The
British, who they were assured were hired bravos engaged by the
Portuguese to capture their deliverers, were proportionately
unpopular, and the movements of Broomfield’s two troops were
again and again betrayed by the natives to the enemy. Often he had
to change his resting-place three and four times a day; he was
engaged with the enemy almost as frequently; yet his active
patrolling continued without interruption, and General Edwards was
kept regularly informed as to every move which the enemy was
making. It was, in its way, an outstanding little bit of work, carried
out with great coolness, persistency and skill, and it by itself would
abundantly have justified all the labour which had been expended in
raising and training the Gold Coast Mounted Infantry.
At Numarroe—which lies much further to the west and must not
be confused with Namirrue—von Lettow-Vorbeck surprised and
captured at the end of August a small British detachment from what
had formerly been one of General Northey’s columns; but at Liome
on the 31st August and on the 1st September he came in for the
worst hammering he had experienced in the whole course of the
campaign, losing some fifty of his Europeans and several hundreds
of his Askari killed, wounded and captured. On this occasion
Lieutenant Drummond and a number of other captives were able to
make their escape.
Thereafter, as is now well known, von Lettow-Vorbeck broke away
north, succeeded in crossing the Lurio River, and thence treked
through the Nyassa Company’s territory to Ngomano on the
Rovuma, where at the end of November in the preceding year he
had re-equipped and refitted at the expense of the Portuguese
garrison. Crossing the Rovuma, he once more entered German East
Africa, still hotly pursued by battalions of the indefatigable King’s
African Rifles; but when after the signing of the Armistice he finally
surrendered, he made his submission to a small police post in
Northern Rhodesia.
The Gold Coast Mounted Infantry, once more reinforced and under
the command of Major Parker, joined in the pursuit as far north as
Ngomano, but on this occasion saw no fighting. On the 3rd October,
however, orders were received for them to return to the Gold Coast
in order to rejoin the 2nd West African Brigade; and as soon as the
necessary arrangements could be completed, the men of the Gold
Coast Mounted Infantry were embarked at Port Amelia, and on their
arrival at Accra were disbanded, and reabsorbed into the Gold Coast
Regiment.
 APPENDIX II

LIST OF HONOURS AND DECORATIONS AWARDED TO EUROPEAN

STAFF AND NATIVE RANK AND FILE OF THE GOLD COAST
REGIMENT DURING THE EAST AFRICAN CAMPAIGN.
(1) European Officers.
Date.
Brevet Lieut.- Temporary Lieut.-Colonel R. A. de 7/2/17
Colonel B. Rose, D.S.O.
Major G. Shaw, M.C. 5/8/18

Brevet Major on Lieut. (Temporary Captain) T. B. C. 5/8/18

promotion to Piggott, M.C.
Captain

D.S.O. Major H. Goodwin 10/6/17

Captain H. A. Harman 10/6/17

Bar to D.S.O. Lieut.-Colonel R. A. de B. Rose, 5/8/18

D.S.O

M.C. Captain (now Lieut.-Colonel) G. 24/11/16

Shaw
Captain A. J. R. O’Brien 24/11/16
Captain R. H. Poyntz 24/1/17
Captain J. Leslie-Smith 13/8/17
 Captain J. G. Foley 29/10/17
Captain H. B. Dawes 5/8/18
Lieutenant T. B. C. Piggott 10/6/17
Lieutenant G. H. Parker 11/3/18
Lieutenant R. F. Beech 11/3/18
Lieutenant G. B. Kinley 30/4/18
Lieutenant L. B. Cumming 27/7/18
Bar to M.C. Captain (now Lieut.-Colonel) G. 13/8/17
Shaw, M.C.
Bar to M.C. Captain A. J. R. O’Brien, M.C. 13/8/17
Captain E. B. Methven, M.C. 5/11/17
Captain J. G. Foley 17/10/18

Legion d’Honneur Lieut.-Colonel R. A. de B. Rose, 22/10/17

Croix d’Officier D.S.O.

Croix de Guerre Major H. Goodwin, D.S.O. 4/1/17

Italian Silver Lieutenant. T. B. C. Piggott, M.C. 4/1/17

Medal

O.B.E. Major H. Read 9/9/18

(2) British Non-commissioned Officers and Men.

D.C.M. 7024 Corpl. J. Campbell 24/1/17

9532 R.S.M. F. C. Medlock 10/6/17
28399 Sergt. E. Thornton 19/7/17
69845 Pte. S. G. Radford 19/7/17
(R.A.M.C.)
1847 Sergt. C. A. Thornett 17/6/18
Bar to D.C.M. 7024 Corpl. J. Campbell 19/7/17

(Russian) Cross 69845 Pte. S. G. Radford 12/11/16

St. George
(3rd Class)

(3) Native Rank and File.

D.C.M 3948 Corpl. Akanno Ibadan 19/7/17

113 M.G.C. John Lagos 19/7/17
3844 C.S.M. Mumuni Moshi 19/7/17
6727 Corpl. Yessufu Kotokoli 19/7/17
5827 Sergt. Moriambah Moshi 19/7/17
5737 Corpl. Musa Fulani 6/7/17
and 19/7/17
6557 Temporary Corpl. Seti 24/11/16
Frafra
8427 Pte. Yaw Kuma 19/7/17
5493 Corpl. and Tem. Sergt. No date
Chililah Grunshi

8581 L/Corpl. Granda Dikale 19/7/17

7339 Trptr. Nuaga Kusase 18/4/17
5048 Corpl. Sandogo Moshi No date
5397 Dr. Musa Karaki No date
5655 Sergt. Alhaji Grunshi 19/7/17
7817 Pte. Seidu Chokosi 20/9/17
5860 L/Corpl. (Acting-Corpl.) 18/10/17
Issaka Dagarti
4188 Sergt. Yessufu Mamprusi 18/10/17
 7426 Bugler Nufu Moshi 1/10/17
4157 C.S.M. Musa Wongara 11/4/17
5225 Sergt. Mamadu Moshi 25/5/17

Bar to D.C.M. 4961 Sergt. Bukara Kukawa 24/11/16

6557 Temporary Corpl. Seti 15/8/17
Frafra

Military Medal 4188 Sergt. Yessufu Mamprusi 19/7/17

6689 Pt. Akuluga Moshi 19/7/17
6414 Sergt. Palpuku Grumah 19/7/17
182 M.G.C. Kwenjeh Moshi 19/7/17
109 M.G.C. Dogali 19/7/17
7842 Pte. Adama Bazaberimi 19/9/17
7248 Pte. Allassan Grumah 15/12/16
4765 Sergt. Braima Dagarti 15/12/16
6690 L/Corpl. Kuka Moshi 15/12/16
6756 Corpl. Timbala Busanga 15/12/16
6675 Corpl. Yero Fulani 15/12/16
13 H.G.C. Imoru Dodo 6/2/17
5593 Corpl. Nuaga Moshi 11/4/17
6688 Pte. Nubela Busanga 11/4/17
6833 Pte. Sebidu Moshi 11/4/17
4388 B.S.M. Bukare Moshi 23/5/17
137 Hdm. G. C. Kwesi John 23/5/17
94 G. C. Lawani Ibadan 23/5/17
959 Sergt. Member 23/5/17
8481 L/Corpl. Ntonge Etun 24/11/16
3851 Sergt. Ali Wongara 24/11/16
170 S. B. Bawa Hausa 24/11/16
200 S. B. Musa Kano 24/11/16
 5658 L/Corpl. Sulley Ibadan 24/1/17

Meritorious V. 103 Corpl. J. W. H. Amartey 17/6/18

Service Medal
O.R.S. G. M. Fraser 17/6/18
31 Qr.Mr.-Sergt. S. Amonoo Aidoo 17/6/18

SERGT. SANDOGO MOSHI, D.C.M.

To face p. 288
 APPENDIX III

STRENGTH OF THE GOLD COAST REGIMENT

ON JULY 31ST, 1916.

Officers 55
British non-commissioned officers 13
Rank and file 1702

STRENGTH OF THE EXPEDITIONARY FORCE ON

JULY 6TH, 1916.

Officers 36
British non-commissioned officers 15
Clerks 11
Rank and File 980
Carriers (battery) 177
Carriers (other) 204
Storemen 1
Officers (R.A.M.C.) 4

STRENGTH OF THE DRAFTS SENT FROM THE GOLD

COAST TO REINFORCE THE EXPEDITIONARY FORCE.

First draft—25th November, 1916:—

 Officers 4
Clerks 1
Rank and file 402
Second draft—21st April, 1917:—
Officers 2
British non-commissioned officers 1
Rank and file 500
Third draft—5th July, 1917:—
Officers 3
British non-commissioned officers 2
Rank and file 799
Fourth draft—6th October, 1917:—
Rank and file 401
Fifth draft—10th December, 1917:—
Rank and file 500
N.B.—Date shown in each case is that of departure from the
Gold Coast.
 APPENDIX IV

Letter from the General Officer Commanding “Pamforce” to

the Officer Commanding the Gold Coast Regiment.

Farewell Message to the Gold Coast Regiment by the

General Officer Commanding Pamforce.
The departure of the Gold Coast Regiment from my Command
furnishes me with a fitting opportunity to place on record my high
appreciation of the distinguished and gallant services which the Gold
Coast Regiment has never failed to render me within the period that
I have had the honour to command Pamforce.
The greatest testimony to the excellence of the services rendered
by the Gold Coast Regiment is to be found in the fact that during the
period which the Regiment has formed an integral part of Pamforce,
it has assisted in reducing the enemy forces by at least one-half of
his former strength, and the measure of the achievement of the
Regiment is the contrast between the strength of the enemy force
when Pamforce was formed and his strength to-day.
I would desire at this juncture to pay a high tribute of my regard
to the Officer Commanding, Officers, British Non-Commissioned
Officers, and the Native Ranks, for the initiative, resource, and
daring which has characterized the service of all during this
particularly difficult phase of the campaign, while I would wish
expressly to place on record my high appreciation and gratitude for
the able and efficient support that has been so loyally extended to
me by Colonel R. A. De B. Rose, D.S.O., to whose soldierly qualities I
feel I owe much.