Professional Documents
Culture Documents
Textbook Intelligent Decision Technologies 2018 Ireneusz Czarnowski Ebook All Chapter PDF
Textbook Intelligent Decision Technologies 2018 Ireneusz Czarnowski Ebook All Chapter PDF
Ireneusz Czarnowski
Visit to download the full and correct content document:
https://textbookfull.com/product/intelligent-decision-technologies-2018-ireneusz-czarn
owski/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...
https://textbookfull.com/product/international-conference-on-
intelligent-data-communication-technologies-and-internet-of-
things-icici-2018-jude-hemanth/
https://textbookfull.com/product/intelligent-decision-technology-
support-in-practice-1st-edition-jeffrey-w-tweedale/
Intelligent Systems for Crisis Management Gi4DM 2018
Orhan Altan
https://textbookfull.com/product/intelligent-systems-for-crisis-
management-gi4dm-2018-orhan-altan/
https://textbookfull.com/product/intelligent-decision-making-in-
quality-management-theory-and-applications-1st-edition-cengiz-
kahraman/
https://textbookfull.com/product/smart-sensors-networks-
communication-technologies-and-intelligent-applications-a-volume-
in-intelligent-data-centric-systems-fatos-xhafa/
Smart Innovation, Systems and Technologies 97
Ireneusz Czarnowski
Robert J. Howlett
Lakhmi C. Jain
Ljubo Vlacic Editors
Intelligent Decision
Technologies 2018
Proceedings of the 10th KES
International Conference on Intelligent
Decision Technologies (KES-IDT 2018)
123
Smart Innovation, Systems and Technologies
Volume 97
Series editors
Robert James Howlett, Bournemouth University and KES International,
Shoreham-by-sea, UK
e-mail: rjhowlett@kesinternational.org
Editors
Intelligent Decision
Technologies 2018
Proceedings of the 10th KES International
Conference on Intelligent Decision
Technologies (KES-IDT 2018)
123
Editors
Ireneusz Czarnowski Lakhmi C. Jain
Gdynia Maritime University Centre for Artificial Intelligence, Faculty of
Gdynia Engineering and Information Technology
Poland University of Technology Sydney
Sydney, NSW
Robert J. Howlett Australia
Bournemouth University and
Poole
UK Faculty of Science, Technology
and Mathematics
and University of Canberra
Canberra, ACT
KES International Australia
Shoreham-by-Sea
UK and
KES International
Shoreham-by-Sea
UK
Ljubo Vlacic
Griffith Sciences - Centres and Institutes
Griffith University
South Brisbane, QLD
Australia
This Springer imprint is published by the registered company Springer International Publishing AG
part of Springer Nature
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
This volume contains the proceedings of the 10th International KES Conference on
Intelligent Decision Technologies (KES-IDT 2018) held in Gold Coast,
Queensland, Australia, on June 20–22, 2018.
KES-IDT is an international annual conference organized by KES International.
The KES-IDT conference is a sub-series of the KES Conference series.
The conference provided opportunities for the presentation of new research
results and discussion about them, leading to knowledge transfer and generation of
new ideas in the field of intelligent decision-making.
This edition, KES-IDT 2018, attracted a number of researchers and practitioners
from all over the world. The KES-IDT 2018 Program Committee accepted 24
papers for oral presentation and publication in the volume of the KES-IDT 2018
proceedings. These papers have been submitted for the main track and special
sessions devoted to specific topics, such as:
• Decision-Making Theory for Economics,
• Advances in Knowledge-based Statistical Data Analysis,
• On Knowledge-Based Digital Ecosystems and Technologies for Smart and
Intelligent Decision Support Systems,
• Soft Computing Models in Industrial and Management Engineering,
• Computational Media Computing and its Applications,
• Intelligent Decision-Making Technologies,
• Digital Architecture and Decision Management.
Each paper has been reviewed by 2–3 members of the International Program
Committee and International Reviewer Board.
We are very satisfied with the quality of the program and would like to thank the
authors for choosing KES-IDT as the forum for the presentation of their work. Also,
we gratefully acknowledge the hard work of the KES-IDT International Program
Committee members and of the additional reviewers for taking the time to review
the submitted papers and selecting the best among them for presentation at the
conference and inclusion in its proceedings.
v
vi Preface
We hope and intend that KES-IDT 2018 significantly contributes to the fulfill-
ment of the academic excellence and leads to even greater successes of KES-IDT
events in the future.
Honorary Chairs
General Chair
Executive Chair
Program Chair
vii
viii KES-IDT 2018 Conference Organization
Publicity Chair
Special Sessions
Decision-Making Theory for Economics
xv
xvi Contents
Aneesha Bakharia(B)
1 Introduction
Non-negative Matrix Factorization (NMF) [12] and Latent Dirichlet Allocation
(LDA) [4] are two (2) popular and useful algorithms that are able to find latent
topics within document collections. While NMF and LDA stem from different
mathematical underpinnings, both algorithms are able to map documents to
topics and words to topics. Figure 1 shows the top words and documents for a
derived topic on “Beginning Teachers”. NMF and LDA are also not hard clus-
tering algorithms as they allow documents and words to be mapped to multiple
topics, a feature that mimics the way, humans, cluster documents. Many users,
c Springer International Publishing AG, part of Springer Nature 2019
I. Czarnowski et al. (Eds.): KES-IDT 2018, SIST 97, pp. 1–10, 2019.
https://doi.org/10.1007/978-3-319-92028-3_1
2 A. Bakharia
particularly if they come from a research background, don’t trust the output of
topic modeling algorithms [2]. The main reasons for the lack of trust are that the
algorithms are perceived to function as black boxes, are unable to support inter-
activity and don’t improve with human feedback. Recently, however, interactive
variants for both NMF [14] and LDA [1] have been proposed that allow domain
knowledge to be provided to the algorithm in the form of specifying topic seed
words (i.e., to create a new topic from a list of words) and identifying topics that
need to either be merged or split.
Interactive variants of NMF [14] and LDA [1] are capable of addressing user
trust concerns. It is rare that when an algorithm is first introduced for the
evaluation to include human participants. Topic modeling algorithms have been
found to be useful in numerous domain applications but in order for their use
to become mainstream particularly by research analysts (including qualitative
content analysts and decision makers), additional research is required to under-
stand how humans interpret topics and whether they are able to effectively use
the built-in algorithmic mechanisms to add their domain knowledge to refine
topics and answer research questions. While studies evaluating interactive topic
modeling algorithms with human participants are starting to be published [3,6],
this important field of research is in its infancy. In this paper, the results of two
(2) experiments involving human participants were analyzed to produce design
guidelines for the development of interactive topic discovery software. Eleven
(11) design guidelines are included in this paper and provide insight into the
functionality required to support human in the loop interaction with topic mod-
eling algorithms.
2 Methodology
Fig. 1. The user interface to display derived topics for Group B participants in Exper-
iment 1.
final rated topics, and the completed questionnaire) led to additional guidelines
for the development of interactive topic modeling algorithms. Participants were
provided with a research question to answer and asked to identify topics that
contained evidence to support a research finding (Fig. 2).
Fig. 2. The user interface designed to support the addition of domain knowledge in
Experiment 2.
The corpus preprocessing steps included stop word removal, word stemming,
and the construction of a bag-of-words matrix (i.e., document-to-term matrix).
The number of topics was included as a parameter that the participants had
to interactively specify via the provided user interface. NMF variant algorithms
Interactive Topic Discovery Systems 5
In this section design guidelines for interactive human in the loop topic discovery
systems are proposed and substantiated with research findings from Experiment
1: Determining Interactivity Requirements and Experiment 2: Evaluating Inter-
active Topic Modeling Algorithms. The design guidelines support research and
decision making processes.
In Experiment 1, participants completing the activity manually (Group A),
frequently mapped a text response to two (2) or more topics thereby employing
a soft-clustering approach. This suggests that algorithms that allow overlapping
cluster membership are closely related to the way humans group documents
together. NMF and LDA, in particular, are not hard clustering approaches. The
importance of using algorithms that support overlap among topics justifies the
first design guideline (Design Guideline 1). Widely used hard clustering algo-
rithms such as k-means are inappropriate as interactive computational aids for
topic discovery because they will not reveal overlap to the user.
Design Guideline 1:
Algorithms used as aids to interactive topic modeling need to discover
and intuitively visualize topic overlap.
Text responses manually placed in finer grained topics were grouped together
by participants (Group A) because they contained a phrase (i.e., multiple words)
or where relationships existed between the words in the documents which were
not directly reflected in the word usage between the documents. A useful form
of interactivity would, therefore, take advantage of the user’s domain knowledge
and allow them to supply sets of related words or phrases or incorporate pre-built
word embeddings [8].
Design Guideline 2:
Algorithms used as aids to interactive topic modeling need to allow the
user to supply sets of related words as domain knowledge and incorporate
pre-built word embeddings in order to produce topics that are similar to
those that could be derived manually by research analysts.
been proposed for the evaluation of topic modeling algorithms. These metrics
have been shown to moderately correlate with human similarity judgments and
could potentially provide a way to communicate topic quality to the user, thereby
preventing researchers and decision makers from reviewing topics that are of poor
quality. These metrics should be displayed with the derived topics to indicate
topic quality, used as a way to rank topics and provide guidance on the selection
of an initial value for k (Design Guideline 3). It should be noted that for the
smaller and more domain-specific corpus that was used in Experiment 1 and 2,
that both the UCI and UMass coherence measures were found not to correlate
with the rating of topics by participants. In the paper introducing the UCI
measure [10], a higher correlation (i.e., 0.7) was achieved for larger corpora with
generic topics whereas for smaller domain-specific corpuses a lower correlation
(i.e., 0.3) was documented. The impact topic overlap has on topic coherence
has also not been evaluated. Additional research studies are therefore needed
to develop coherence measures that match the human evaluation of overlapping
topics derived from smaller domain-specific corpora.
Design Guideline 3:
Display metrics that indicate the topic coherence of a derived topic to
help the user identify topic quality, rank topics and select an optimal
number of topics.
The Experiment 1 user interface, displayed topics with their top words and
text responses. A weighting was included for both the top words and docu-
ments in a topic. Participants were able to click on a word and view where the
word occurred (i.e., keywords-in-context functionality). Numerous participants
found it difficult to interpret the words in isolation and thought that multi-word
expressions and phrases would be more useful. The data pre-processing sepa-
rated the text responses into unigrams (i.e., single words for inclusion in the
term-document matrix). An n-gram or equivalent approach that is able to iden-
tify phrases would be preferred by qualitative researchers and decision makers
because multi-word expressions and phrases improve the interpretation of topics
(Design Guideline 4). Additional algorithms can also be used such as a topic
labeling technique [7].
Design Guideline 4:
Include phrases or multi-word expressions as features in the bag-of-words
model to allow the top weighted phrases or multi-word expressions in a
topic to be displayed in order to improve topic interpretation.
Fig. 3. Group A participants selection of k and the types of rules specified in Experi-
ment 2.
Design Guideline 5:
Use parameter initialization techniques to aid users with parameter selec-
tion (e.g., specifying the number of topics) for topic modeling algorithms.
It was evident that the display of the top-weighted words in a topic played
a role in helping participants to interpret topics and justifies the inclusion of
Design Guideline 6. Displaying the raw scores from the term-topic matrix next
to a word, however, was found to confuse participants. An alternate technique to
express the raw word scores from NMF as a probability needs to be investigated.
Design Guideline 6:
Allow users to view the top weighted words from a topic as a means by
which to interpret topics at a high-level.
8 A. Bakharia
Allowing participants to click on the top weighted words and view where they
were used within the documents, assisted participants with topic interpretation.
The inclusion of tools that facilitate in-context analysis such as keyword-in-
context tools also led to improved topic interpretation (Design Guideline 7).
The inclusion of keywords-in-context tools facilitated an enhanced understanding
of the topic and provided validating evidence. Current tools that display the
output of topic modeling algorithms [13] don’t show the top documents in a
topic or allow the user to view where the top words have appeared within text
snippets. Findings from the two experiments conducted, found the inclusion
of keyword-in-context functionality crucial to support topic interpretation and
evidence gathering.
Design Guideline 7:
The inclusion of keyword-in-context tools is important to facilitate topic
interpretation and support evidence gathering.
Design Guideline 8:
Interactive topic modeling algorithms need to allow users to specify new
topics (i.e., topic seeding), merge topics and split topics.
Design Guideline 9:
Interactivity should not only be one-way between the user and the algo-
rithm, models of interactivity based on Active Machine Learning and
word embedding techniques must be leveraged to facilitate algorithm-
user interactivity.
Interactive Topic Discovery Systems 9
In this paper, the findings of two (2) experiments using both generic and inter-
active variants of NMF and LDA were used to derive design guidelines for inter-
active topic discovery software systems. The proposed guidelines are practical
and indicate that a holistic approach to the design and implementation of topic
discovery software are essential when humans need to answer specific research
questions and make decisions based on their findings. In particular, the inclusion
of numerous algorithms in addition to the interactive variants of NMF and LDA
are required. Key additional algorithms include, algorithms to aid in the selection
of the number of topics k, algorithms that are able to calculate topic coherence
on domain-specific corpora, active learning algorithms to enable two-way initi-
ated interactivity (i.e., human-algorithm and algorithm-human), and similarity
metrics to help the user find topics that have emerged as a result of user-supplied
domain knowledge (i.e., seed words to create topics, topic merge rules, and topic
split rules). Intuitive user interface design is also required to assist the user
in specifying domain knowledge. Future research directions include developing
and evaluating an open source system that incorporates the design guidelines
outlined in this paper.
References
1. Andrzejewski, D., Zhu, X., Craven, M.: Incorporating domain knowledge into topic
modeling via dirichlet forest priors. In: Proceedings of the 26th Annual Interna-
tional Conference on Machine Learning, pp. 25–32. ACM (2009)
2. Bakharia, A.: Interactive content analysis: evaluating interactive variants of non-
negative Matrix Factorisation and Latent Dirichlet Allocation as qualitative con-
tent analysis aids. Ph.D. thesis, Queensland University of Technology (2014)
3. Bakharia, A., Bruza, P., Watters, J., Narayan, B., Sitbon, L.: Interactive topic
modeling for aiding qualitative content analysis. In: Proceedings of the 2016 ACM
on Conference on Human Information Interaction and Retrieval, pp. 213–222. ACM
(2016)
4. Blei, D.M., Ng, A.Y., Jordan, M.I.: Latent dirichlet allocation. J. Mach. Learn.
Res. 3, 993–1022 (2003)
5. Boutsidis, C., Gallopoulos, E.: SVD based initialization: a head start for nonneg-
ative matrix factorization. Pattern Recogn. 41(4), 1350–1362 (2008)
6. Lee, T.Y., Smith, A., Seppi, K., Elmqvist, N., Boyd-Graber, J., Findlater, L.: The
human touch: how non-expert users perceive, interpret, and fix topic models. Int.
J. Hum Comput Stud. 105, 28–42 (2017)
7. Mei, Q., Shen, X., Zhai, C.: Automatic labeling of multinomial topic models. In:
Proceedings of the 13th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, pp. 490–499. ACM (2007)
8. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed repre-
sentations of words and phrases and their compositionality. In: Advances in Neural
Information Processing Systems, pp. 3111–3119 (2013)
9. Mimno, D., Wallach, H.M., Talley, E., Leenders, M., McCallum, A.: Optimizing
semantic coherence in topic models. In: Proceedings of the Conference on Empirical
Methods in Natural Language Processing, pp. 262–272. Association for Computa-
tional Linguistics (2011)
10. Newman, D., Noh, Y., Talley, E., Karimi, S., Baldwin, T.: Evaluating topic models
for digital libraries. In: Proceedings of the 10th Annual Joint Conference on Digital
Libraries, pp. 215–224. ACM (2010)
11. O’Callaghan, D., Greene, D., Carthy, J., Cunningham, P.: An analysis of the coher-
ence of descriptors in topic modeling. Expert Syst. Appl. 42(13), 5645–5657 (2015)
12. Seung, D., Lee, L.: Algorithms for non-negative matrix factorization. Adv. Neural
Inf. Proc. Syst. 13, 556–562 (2001)
13. Sievert, C., Shirley, K.E.: LDAvis: A method for visualizing and interpreting topics
(2014)
14. Wang, F., Li, T., Zhang, C.: Semi-supervised clustering via matrix factorization.
In: SDM, pp. 1–12 (2008)
Daily Stress Recognition System Using
Activity Tracker and Smartphone Based
on Physical Activity and Heart Rate Data
1 Introduction
Daily stress has been a topic for a long time. Despite there are many factors
related to stress, the consequences of stress obviously affect the quality of life
and wellbeing. A study suggested that physical symptoms are likely to occur
after a stressful day [3]. In general, stress management is an important issue and
it is necessary to make use of daily technology and devices to help improve one’s
wellbeing as much as possible.
In this paper, we proposed a system that is based on Internet of Things (IoT)
architecture to recognize stress level in daily life. The system adopted an activity
tracker as a sensing device. The activity tracker itself is also capable of sensing
c Springer International Publishing AG, part of Springer Nature 2019
I. Czarnowski et al. (Eds.): KES-IDT 2018, SIST 97, pp. 11–21, 2019.
https://doi.org/10.1007/978-3-319-92028-3_2
12 W. Lawanont et al.
heart rate, which is an important feature for stress recognition [15]. The system
was developed based on the idea that it must be practical for daily usage and it
should work with devices that can be easily found in the market.
2 Related Work
After the emerging of technology, countless of studies were conducted to improve
one’s wellbeing. These studies include several perspectives, one is the study con-
ducted in clinical setting to help the patient with chronic diseases or similar
cases, while the other studies try to improve the wellbeing by using daily life tech-
nology to solve the matter [8,17]. The activity recognition was further used for
stress recognition as well. Nicholas and Jean-Marc proposed a study to develop a
supervised learning method for recognizing stress [11]. The study collected phys-
iological data in a stressful situation that was built for the experiment. Beside
from physiological data classification model, a study discussed stress recognition
using smartphone data, weather conditions, and individual traits [1]. On the
other hand, a study called StressSense used smartphone to recognize the stress
[9]. These two studies demonstrated good examples of using a simple everyday
device, such as smartphone, to help improve one’s wellbeing.
As an extension to smartphone data, such as number of calls and physical
activities, another parameter that was proposed to help recognize the stress
was heart rate variability (HRV) [15]. The study findings suggested that the
HRV may index important organism functions and it should be considered as
a potential stress marker. With the use of heart rate approach, a study used a
low cost chest strap heart rate sensor to measure stress [13]. Despite that the
result was very satisfying, the heart rate chest strap is still not widely used. On
another perspective, a study by Muaremi et al. showed an example of measuring
stress using the combination of smartphone and wearable devices [12]. However,
the wearable device used in the study is also not widely used.
In general, the approach of using smartphone and wearable devices were
adopted for stress recognition. However, the works which were feasible for daily
usage did not use heart rate or HRV as one of the parameter for model train-
ing. In addition, the user was required to wear a special device, such as chest
strap, for heart rate monitoring. Thus, we proposed a solution where the system
used smartphone and activity tracker for stress recognition while consider both
physical activities data as well as heart rate data from wearable device.
3 Methodology
3.1 System Overview
Figure 1 shows the overview of the proposed system. This study designed the
system according to the ETSI M2M architecture [6]. In this system, we divided
the system into two main parts. The body area network consists of two main
components, activity tracker and smartphone. The activity tracker serves as
Another random document with
no related content on Scribd:
deux ailes d'infanterie devaient se déployer, et envelopper l'armée
de Constantin, déja rompue par la cavalerie. Le prince, qui avait le
coup d'œil militaire, comprit le dessein des ennemis à l'ordre de leur
bataille. Il place des corps à droite et à gauche pour faire face à
l'infanterie et arrêter ses mouvements. Pour lui, il se met au centre
en tête de cette redoutable cavalerie. Quand il la voit sur le point de
heurter le front de son armée, au lieu de lui résister, il ordonne à ses
troupes de s'ouvrir: c'était un torrent qui n'avait de force qu'en ligne
droite; le fer dont elle était revêtue ôtait toute souplesse aux
hommes et aux chevaux. Mais dès qu'il la voit engagée entre ses
escadrons, il la fait enfermer et attaquer de toutes parts, non pas à
coups de lances et d'épée, on ne pouvait percer de tels ennemis,
mais à grands coups de masses d'armes. On les assommait, on les
écrasait sur la selle de leurs chevaux, on les renversait, sans qu'ils
pussent ni se mouvoir pour se défendre, ni se relever quand ils
étaient abattus. Bientôt ce ne fut plus qu'une horrible confusion
d'hommes, de chevaux, d'armes, amoncelés les uns sur les autres.
Ceux qui échappèrent à ce massacre voulurent se sauver à Turin
avec l'infanterie: mais ils en trouvèrent les portes fermées, et
Constantin, qui les poursuivit l'épée dans les reins, acheva de les
tailler en pièces au pied des murailles.
Cette victoire, qui ne coûta point de sang au
vainqueur, lui ouvrit les portes de Turin. La plupart iv. Suites de la
des autres places entre le Pô et les Alpes lui victoire.
envoyèrent des députés pour l'assurer de leur
soumission: toutes s'empressaient de lui offrir des Incert. Pan. c. 7.
vivres. Sigonius, sur un passage de saint Jérôme,
conjecture que Verceil fit quelque résistance, et
que cette ville fut alors presque détruite. Il n'en est Sigon. Imp.
Occ. p. 52.
point parlé ailleurs. Constantin alla à Milan, et son
entrée devint une espèce de triomphe par la joie et
les acclamations des habitants, qui ne pouvaient Hieron. Epist.
se lasser de le voir et de lui applaudir comme au ad Innocentium.
libérateur de l'Italie. t. i, p. 3. ed.
Vallars.
Au sortir de Milan, où il était resté quelques jours
pour donner du repos à ses troupes, il prit la route de Vérone. Il
savait qu'il y trouverait rassemblées les plus
grandes forces de Maxence, commandées par les v. Siége de
meilleurs capitaines de ce prince et par son préfet Vérone.
du prétoire, Ruricius Pompéianus, le plus brave et
le plus habile général que le tyran eût à son Incert. Pan. c. 8
service. En passant auprès de Bresce [Brixia], et seq.
Constantin rencontra un gros corps de cavalerie,
qui prit la fuite au premier choc et alla rejoindre
Nazar. Pan. c.
l'armée de Vérone. Ruricius n'osa tenir la 26.
campagne; il se renferma avec ses troupes dans la
ville. Le siége en était difficile: il fallait passer l'Adige [Athesis], et se
rendre maître du cours de ce fleuve qui portait l'abondance à
Vérone: il était rapide, plein de gouffres et de rochers, et les ennemis
en gardaient les bords. Constantin trompa pourtant leur vigilance;
étant remonté fort au-dessus de la ville, jusqu'à un endroit où le
trajet était praticable, il y fit passer à leur insu une partie de son
armée. A peine le siége fut-il formé, que les assiégés firent une
vigoureuse sortie, et furent repoussés avec tant de carnage, que
Ruricius se vit obligé de sortir secrètement de la ville pour aller
chercher de nouveaux secours.
Il revint bientôt avec une plus grosse armée, résolu
de faire lever le siége ou de périr. L'empereur, pour vi. Bataille de
ne pas donner aux assiégés la liberté de Vérone.
s'échapper, ou même de l'attaquer en queue
pendant le combat, laisse devant la ville une partie Incert. Pan. c. 9
de ses troupes, et marche avec l'autre à la et 10.
rencontre de Ruricius. Il range d'abord son armée
sur deux lignes; mais ayant observé que celle des
Nazar. Pan. c.
ennemis était plus nombreuse, il met la sienne sur 26.
une seule ligne, et fait un grand front de peur d'être
enveloppé. Le combat commença sur le déclin du jour, et dura fort
avant dans la nuit. Constantin y fit le devoir de général et de soldat.
Il se jette au plus fort de la mêlée, et profitant des ténèbres pour
courir, sans être retenu, où l'emportait sa valeur, il perce, il abat, il
terrasse; on ne le reconnaît qu'à la pesanteur de son bras: le son
des instruments de guerre, le cri des soldats, le cliquetis des armes,
les gémissements des blessés, les coups guidés par le hasard, tant
d'horreurs augmentées par celle d'une nuit épaisse, ne troublent
point son courage. L'armée de secours est entièrement défaite;
Ruricius y perd la vie: Constantin hors d'haleine, couvert de sang et
de poussière, va rejoindre les troupes du siége, et reçoit de ses
principaux officiers, qui s'empressent avec des larmes de joie de
baiser ses mains sanglantes, des reproches d'autant plus flatteurs,
qu'ils sont mieux mérités.
Pendant le siége de Vérone, Aquilée [Aquileia] et
Modène [Mutina] furent attaquées: elles se vii. Prise de
rendirent, avec plusieurs autres villes, en même Vérone,
temps que Vérone. L'empereur accorda la vie aux d'Aquilée et de
Modène.
habitants; mais il les obligea de rendre leurs
armes; et pour s'assurer de leurs personnes, il les
mit sous la garde de ses soldats. Comme ils Incert. Pan. c.
étaient en plus grand nombre que les vainqueurs, 11 et seq.
on crut nécessaire de les enchaîner, et on
manquait de chaînes; Constantin leur en fit faire de Nazar. c. 27.
leurs propres épées, qui, forgées pour leur
défense, devinrent les instruments de leur servitude.
Après tant d'heureux succès, rien n'arrêta sa
marche jusqu'à la vue de Rome. Il paraît viii. Constantin
seulement par un mot de Lactance, qu'aux devant Rome.
approches de cette ville il éprouva quelque revers;
mais que sans perdre courage, et déterminé à tout Lact. de mort.
événement, il marcha en avant et vint camper vis- persec. c. 44.
à-vis du Ponte-Mole, nommé alors le pont Milvius.
C'est un pont de pierre de huit arches sur le Tibre,
à deux milles au-dessus de Rome dans la voie Fabric. descript.
urb. Rom. c. 16,
Flaminia, par laquelle venait Constantin. Il avait été et alii passim.
construit en bois dès les premiers siècles de la
république; il fut rebâti en pierres par le censeur Emilius Scaurus, et
rétabli par Auguste. Il subsiste encore aujourd'hui, ayant été réparé
par le pape Nicolas V, au milieu du quinzième siècle.
Tout ce que craignait Constantin, c'était d'être
obligé d'assiéger Rome, bien pourvue de troupes ix. Maxence se
et de toutes sortes de munitions; et de faire tient enfermé
ressentir les calamités de la guerre à un peuple dans Rome.
dont il voulait se faire aimer. Maxence, soit par
lâcheté, soit par une crainte superstitieuse, se Incert. Pan. c.
tenait renfermé: on lui avait prédit qu'il périrait, s'il 14 et seq.
sortait hors des portes de la ville; il n'osait même
quitter son palais, que pour se transporter aux
Lact., de mort.
jardins délicieux de Salluste. Cependant affectant pers. c. 44.
une fausse confiance, il n'avait rien retranché de
ses débauches ordinaires. Par une précaution
frivole, il avait supprimé toutes les lettres qui Noris. in num.
annonçaient ses infortunes; il supposait même des Diocl. c. 5.
victoires pour amuser le peuple: et ce fut
apparemment dans ce temps-là qu'il se fit décorer tant de fois du
titre d'Imperator, qui lui est donné pour la onzième fois sur un
marbre antique; vanité ridicule, qui donne à la postérité, plus
exactement que l'histoire même, le calcul de ses pertes. Quelquefois
il protestait hautement que tous ses désirs étaient de voir son rival
au pied des murs de Rome, se flattant sans doute de lui débaucher
son armée, et peu capable de sentir la différence qu'il devait y avoir
entre les troupes de Sévère ou de Galérius, et des soldats conduits
par Constantin et par la victoire. Il s'en fallait bien qu'il fût aussi
tranquille, qu'il affectait de le paraître. Deux jours avant la bataille,
effrayé par des présages et par des songes, que sa timidité
interprétait d'une manière funeste, il quitta son palais, et alla s'établir
avec sa femme et ses enfants dans une maison particulière.
Cependant son armée sortit de Rome, et se posta vis-à-vis de celle
de Constantin, le Ponte-Mole entre deux.
Ce dut être alors que Maxence fit jeter un pont de
bateaux sur le fleuve, au-dessus du Ponte-Mole, x. Pont de
apparemment vers l'endroit appelé les Roches- bateaux.
Rouges [Saxa rubra], à neuf milles de Rome.
C'était le lieu qu'il avait choisi pour combattre, soit Euseb. vit.
que le poste lui parût plus avantageux, soit pour Const. l. 1, c.
obliger ses troupes à faire de plus grands efforts 38.
en leur rendant la retraite plus difficile, soit que, se
défiant des Romains, il voulût livrer la bataille hors Zos. l. 2, c. 15.
de leur vue. Ce pont était construit de manière qu'il
pouvait s'ouvrir ou se rompre en un moment,
n'étant lié par le milieu qu'avec des crampons de Aurel. Vict. de
fer, qu'il était aisé de détacher. C'était en cas de Cæs. p. 175.
défaite un moyen de faire périr l'armée victorieuse
dans le temps même de la poursuite. Des ouvriers Vict. epit. p.
cachés dans les bateaux devaient ouvrir le pont, 221.
dès que Constantin et ses troupes seraient
dessus, pour les précipiter dans le fleuve.
Lact. de mort.
Quelques modernes, fondés sur le récit que persec. c. 44.
Lactance, les panégyristes et Prudence font de
cette bataille, nient l'existence de ce pont; ils
prétendent que ce fut du pont Milvius que Libanius, orat.
Maxence dans sa déroute tomba dans le Tibre, 3. t. 11, p. 105
et 106. ed.
soit qu'il l'eût lui-même fait rompre avant l'action, Morel.
comme Lactance semble le dire, soit que la foule
des fuyards l'en ait précipité. Mais nous suivrons
ici Eusèbe et Zosime, qui décrivent en termes Praxag. apud
précis ce pont de bateaux, et dont le témoignage Phot. cod. 62.
très-considérable en lui-même, surtout quand ils
s'accordent ensemble, est ici appuyé par le plus Acta Metr. et
grand nombre d'anciens auteurs. Alex. apud
Phot. cod. 256.
La nuit qui précéda la bataille, Constantin fut averti
en songe de faire marquer les boucliers de ses
soldats du monogramme de Christ. Il obéit, et dès Incert. Pan. c.
le point du jour ce victorieux caractère, imprimé 27.
par son ordre, parut sur les boucliers, sur les
casques, et fit passer dans le cœur des soldats Prud. ad Sym. l.
une confiance toute nouvelle. 1, va. 448.