Agents and Data Mining Interaction

10th International Workshop ADMI 2014

Paris France May 5 9 2014 Revised
Selected Papers 1st Edition Longbing
Cao
Visit to download the full and correct content document:
https://textbookfull.com/product/agents-and-data-mining-interaction-10th-international
-workshop-admi-2014-paris-france-may-5-9-2014-revised-selected-papers-1st-edition
-longbing-cao/
More products digital (pdf, epub, mobi) instant
download maybe you interests ...

Engineering Multi Agent Systems Second International

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

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

Multi Agent Based Simulation XV International Workshop

MABS 2014 Paris France May 5 6 2014 Revised Selected
Papers 1st Edition Francisco Grimaldo

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

Constructive Side Channel Analysis and Secure Design

5th International Workshop COSADE 2014 Paris France
April 13 15 2014 Revised Selected Papers 1st Edition
Emmanuel Prouff (Eds.)
https://textbookfull.com/product/constructive-side-channel-
analysis-and-secure-design-5th-international-workshop-
cosade-2014-paris-france-april-13-15-2014-revised-selected-
papers-1st-edition-emmanuel-prouff-eds/

Modelling and Simulation for Autonomous Systems First

International Workshop MESAS 2014 Rome Italy May 5 6
2014 Revised Selected Papers 1st Edition Jan Hodicky
(Eds.)
https://textbookfull.com/product/modelling-and-simulation-for-
autonomous-systems-first-international-workshop-mesas-2014-rome-
italy-may-5-6-2014-revised-selected-papers-1st-edition-jan-
Rewriting Logic and Its Applications 10th International
Workshop WRLA 2014 Held as a Satellite Event of ETAPS
Grenoble France April 5 6 2014 Revised Selected Papers
1st Edition Santiago Escobar (Eds.)
https://textbookfull.com/product/rewriting-logic-and-its-
applications-10th-international-workshop-wrla-2014-held-as-a-
satellite-event-of-etaps-grenoble-france-april-5-6-2014-revised-
selected-papers-1st-edition-santiago-escobar-eds/

Multimodal Analyses enabling Artificial Agents in Human

Machine Interaction Second International Workshop
MA3HMI 2014 Held in Conjunction with INTERSPEECH 2014
Singapore Singapore September 14 2014 Revised Selected
Papers 1st Edition Ronald Böck
https://textbookfull.com/product/multimodal-analyses-enabling-
artificial-agents-in-human-machine-interaction-second-
international-workshop-ma3hmi-2014-held-in-conjunction-with-
interspeech-2014-singapore-singapore-september-14-2014-re/

Swarm Intelligence Based Optimization First

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

Semantic Web Collaborative Spaces Second International

Workshop SWCS 2013 Montpellier France May 27 2013 Third
International Workshop SWCS 2014 Trentino Italy October
19 2014 Revised Selected and Invited Papers 1st Edition
Pascal Molli
https://textbookfull.com/product/semantic-web-collaborative-
spaces-second-international-workshop-swcs-2013-montpellier-
france-may-27-2013-third-international-workshop-
swcs-2014-trentino-italy-october-19-2014-revised-selected-and-
invi/
Provenance and Annotation of Data and Processes 5th
International Provenance and Annotation Workshop IPAW
2014 Cologne Germany June 9 13 2014 Revised Selected
Papers 1st Edition Bertram Ludäscher
https://textbookfull.com/product/provenance-and-annotation-of-
data-and-processes-5th-international-provenance-and-annotation-
workshop-ipaw-2014-cologne-germany-june-9-13-2014-revised-
 Longbing Cao · Yifeng Zeng
Bo An · Andreas L. Symeonidis
Vladimir Gorodetsky · Frans Coenen
Philip S. Yu (Eds.)

Agents
LNAI 9145

and Data Mining

Interaction
10th International Workshop, ADMI 2014
Paris, France, May 5–9, 2014
Revised Selected Papers

123
Lecture Notes in Artificial Intelligence 9145

Subseries of Lecture Notes in Computer Science

LNAI Series Editors

Randy Goebel
University of Alberta, Edmonton, Canada
Yuzuru Tanaka
Hokkaido University, Sapporo, Japan
Wolfgang Wahlster
DFKI and Saarland University, Saarbrücken, Germany

LNAI Founding Series Editor

Joerg Siekmann
DFKI and Saarland University, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/1244
Longbing Cao Yifeng Zeng
•

Bo An Andreas L. Symeonidis
•

Vladimir Gorodetsky Frans Coenen

•

Philip S. Yu (Eds.)

Agents
and Data Mining
Interaction
10th International Workshop, ADMI 2014
Paris, France, May 5–9, 2014
Revised Selected Papers

123
Editors
Longbing Cao Vladimir Gorodetsky
University of Technology Sydney Russian Academy of Sciences
Sydney, NSW St. Petersburg
Australia Russia
Yifeng Zeng Frans Coenen
Teesside University University of Liverpool
Middlesborough Liverpool
UK UK
Bo An Philip S. Yu
Nanyang Technological University University of Illinois at Chicago
Singapore Chicago, IL
Singapore USA
Andreas L. Symeonidis
Aristotle University of Thessaloniki
Thessaloniki
Greece

ISSN 0302-9743 ISSN 1611-3349 (electronic)

Lecture Notes in Artificial Intelligence
ISBN 978-3-319-20229-7 ISBN 978-3-319-20230-3 (eBook)
DOI 10.1007/978-3-319-20230-3

Library of Congress Control Number: 2015940986

LNCS Sublibrary: SL7 – Artificial Intelligence

Springer Cham Heidelberg New York Dordrecht London

© Springer International Publishing Switzerland 2015
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the
material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now
known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book are
believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors
give a warranty, express or implied, with respect to the material contained herein or for any errors or
omissions that may have been made.

Printed on acid-free paper

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

(www.springer.com)
 Message from the Workshop Chairs

We are pleased to welcome you to the proceedings of the 2014 International Workshop
on Agents and Data Mining Interaction (ADMI 2014), held jointly with AAMAS 2014.
In recent years, agents and data mining interaction (ADMI, or agent mining) has
emerged as a very promising research field. Following the success of previous ADMIs,
ADMI 2014 provided a premier forum for sharing research and engineering results, as
well as potential challenges and prospects encountered in the coupling between agents
and data mining.
The ADMI 2014 workshop encouraged and promoted theoretical and applied
research and development, which aims at:
– Exploiting agent-enriched data mining and demonstrating how intelligent agent
technology can contribute to critical data mining problems in theory and practice
– Improving data mining-driven agents and showing how data mining can strengthen
agent intelligence in research and practical applications
– Exploring the integration of agents and data mining toward a super-intelligent
system
– Discussing existing results, new problems, challenges and the impact of the inte-
gration of agent and data mining technologies as applied to highly distributed
heterogeneous, including mobile, systems operating in ubiquitous and P2P
environments
– Identifying challenges and directions for future research and development on the
synergy between agents and data mining
The 10 papers included in this edition are from eight countries. ADMI 2014 sub-
missions represented areas from America, Europe, to Asia, indicating the boom of
agent mining research globally. The workshop also included two invited talks by two
distinguished researchers.
As with previous ADMIs, the papers accepted for ADMI 2014 were revised and are
included in this LNAI proceedings volume published by Springer. We appreciate
Springer, in particular Alfred Hofmann, for the continuing publication support.
ADMI 2014 was sponsored by the Special Interest Group: Agent-Mining Interaction
and Integration (AMII-SIG: www.agentmining.org). We appreciate the guidance of the
Steering Committee.
More information about ADMI 2014 is available from the workshop website: http://
admi14.agentmining.org/.
VI Message from the Workshop Chairs

Finally, we appreciate the contributions made by all authors, Program Committee

members, invited speakers, panelists, and AAMAS 2014 workshop and local
organizers.

May 2014 Philip S. Yu

Longbing Cao
Yifeng Zeng
Bo An
Andreas L. Symeonidis
Vladimir Gorodetsky
Frans Coenen
 Organization

General Chair
Philip S. Yu University of Illinois at Chicago, USA

Workshop Co-chairs
Longbing Cao University of Technology Sydney, Australia
Yifeng Zeng Teesside University, UK
Bo An Nanyang Technological University, Singapore
Andreas L. Symeonidis Aristotle University of Thessaloniki, Greece
Vladimir Gorodetsky Russian Academy of Sciences, Russia
Frans Coenen University of Liverpool, UK

Workshop Organizing Co-chair

Ross Conroy Teesside University, UK
Hua Mao Sichuan University, China
Yingke Chen University of Georgia, USA

Program Committee
Ahmed Hambaba San Jose State University, USA
Ajith Abraham Norwegian University of Science and Technology,
Norway
Andrea G.B. Tettamanzi University of Milan, Italy
Andreas L. Symeonidis Aristotle University of Thessaloniki, Greece
Andrzej Skowron Institute of Decision Process Support, Poland
Bo Zhang Tsinghua University, China
Daniel Kudenko University of York, UK
Daniel Zeng Arizona University, USA
David Taniar Monash University, Australia
Deborah Richards Macquarie University, Australia
Dionysis Kehagias Informatics and Telematics Institute, Greece
Eduardo Alonso University of York, UK
Eugenio Oliveira University of Porto, Portugal
Frans Oliehoek Massachusetts Institute of Technology, USA
Gao Cong Nanyang Technological University, Singapore
Henry Hexmoor University of Arkansas, USA
Ioannis Athanasiadis Democritus University of Thrace, Greece
Jason Jung Yeungnam University, Korea
VIII Organization

Joerg Mueller Technische Universität Clausthal, Germany

Juan Carlos Cubero University of Granada, Spain
Katia Sycara Carnegie Mellon University, USA
Kazuhiro Kuwabara Ritsumeikan University, Japan
Kim-leng Poh National University of Singapore
Leonid Perlovsky AFRL/IFGA, USA
Luis Otavio Alvares Universidade Federal do Rio Grande do Sul, Brazil
Martin Purvis University of Otago, New Zealand
Michal Pechoucek Czech Technical University, Czech Republic
Mingyu Guo University of Liverpool, UK
Nathan Griffths University of Warwick, UK
Pericles A. Mitkas Aristotle University of Thessaloniki, Greece
Ran Wolff Haifa University, Israel
Seunghyun Im University of Pittsburgh at Johnstown, USA
Stefan Witwicki Instituto Superior Técnico, Portugal
Sung-Bae Cho Yonsei University, Korea
Sviatoslav Braynov University of Illinois at Springfield, USA
Tapio Elomaa Tampere University of Technology, Finland
Valerie Camps Paul Sabatier University, France
Vladimir Gorodetsky SPIIRAS, Russia
Wen-Ran Zhang Georgia Southern University, USA
William Cheung Hong Kong Baptist University, SAR China
Xudong Luo Sun Yat-sen University, China
Yan Wang Macquarie University, Australia
Yinghui Pan Jiangxi University of Finance and Economics, China
Yves Demazeau CNRS, France
Zbigniew Ras University of North Carolina, USA
Zili Zhang Deakin University, Australia
Zinovi Rabinovich University of Southampton, UK

Steering Committee
Longbing Cao University of Technology Sydney, Australia (Coordinator)
Edmund H. Durfee University of Michigan, USA
Vladimir Gorodetsky St. Petersburg Institute for Informatics and Automation,
Russia
Hillol Kargupta University of Maryland Baltimore County, USA
Matthias Klusch DFKI, Germany
Michael Luck King’s College London, UK
Jiming Liu Hong Kong Baptist University, SAR China
Pericles A. Mitkas Aristotle University of Thessaloniki, Greece
Joerg Mueller Technische Universität Clausthal, Germany
Ngoc Thanh Nguyen Wroclaw University of Technology, Poland
Carles Sierra Artificial Intelligence Research Institute of the Spanish
Research Council, Spain
Andreas L. Symeonidis Aristotle University of Thessaloniki, Greece
 Organization IX

Gerhard Weiss University of Maastricht, The Netherlands

Xindong Wu University of Vermont, USA
Philip S. Yu University of Illinois at Chicago, USA
Chengqi Zhang University of Technology Sydney, Australia
 Contents

Learning Agents’ Relations in Interactive Multiagent Dynamic Influence

Diagrams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
Yinghui Pan, Yifeng Zeng, and Hua Mao

Agent-Based Customer Profile Learning in 3G Recommender Systems:

Ontology-Driven Multi-source Cross-Domain Case . . . . . . . . . . . . . . . . . . . 12
Vladimir Gorodetsky, Vladimir Samoylov, and Olga Tushkanova

Modeling Temporal Propagation Dynamics in Multiplex Networks . . . . . . . . 26

Zhaofeng Li, Fuhan Yan, and Yichuan Jiang

Mining Movement Patterns from Video Data to Inform Multi-agent Based

Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Muhammad Tufail, Frans Coenen, Tintin Mu, and Saqib Jamshid Rind

Accessory-Based Multi-agent Simulating Platform on the Web . . . . . . . . . . . 52

Jinyu Zhang, Chenhui Xia, and Weibo Zhang

Performance Evaluation of Agents and Multi-agent Systems Using Formal

Specifications in Z Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
Christos Dimou, Fani Tzima, Andreas L. Symeonidis,
and Pericles A. Mitkas

Reputation in Communities of Agent-Based Web Services Through

Data Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Mohamad Mehdi, Nizar Bouguila, and Jamal Bentahar

Data Mining Process Optimization in Computational Multi-agent Systems . . . 93

Ondřej Kazík and Roman Neruda

Diversifying the Storytelling Using Bayesian Networks . . . . . . . . . . . . . . . . 104

Wenyun Wu, Biyang Ma, Shaoxin Zhang, Yifeng Zeng, and Hua Mao

A Coupled Similarity Kernel for Pairwise Support Vector Machine . . . . . . . . 114

Mu Li, Jinjiu Li, Yuming Ou, and Dan Luo

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

 Learning Agents’ Relations in Interactive
Multiagent Dynamic Influence Diagrams

Yinghui Pan1,2,3,4 , Yifeng Zeng2(B) , and Hua Mao1,2,3,4

1
School of Information Technology,
Jiangxi University of Finance and Economics, Nanchang, China
panyinghui@jxufe.edu.cn
2
Department of Automation, Xiamen University, Xiamen, China
yfzeng@xmu.edu.cn
3
School of Computing, Teesside University, Middlesbrough, UK
4
College of Computer Science, Sichuan University, Chengdu, China

Abstract. Solving interactive multiagent decision making problems is

a challenging task since it needs to model how agents interact over time.
From individual agents’ perspective, interactive dynamic influence dia-
grams (I-DIDs) provide a general framework for sequential multiagent
decision making in uncertain settings. Most of the current I-DID research
focuses on the setting of n = 2 agents, which limits its general applica-
tions. This paper extends I-DIDs for n > 2 agents, which as expected
increases the solution complexity due to the model space of other agents
in the extended I-DIDs. We exploit data of agents’ interactions to dis-
cover their relations thereby reducing the model complexity. We show
preliminary results of the proposed techniques in one problem domain.

Keywords: Interactive dynamic influence diagrams · Relation learn-

ing · Intelligent agents

1 Introduction
Extending single-agent based graphical models of influence diagrams [7], interac-
tive dynamic influence diagrams (I-DIDs) [4,16] provide a general framework for
solving sequential multiagent decision making problem under uncertainty. Differ-
ing from other frameworks such as Dec-POMDPs [12] and multiagent influence
diagrams [10], I-DIDs solve the problem from the perspective of an individual
agent and do not make the common belief assumption in modeling other agents.
Hence I-DIDs are a general decision model and may be employed to solve both
cooperative and competitive multiagent decision problems.
Algorithms for solving I-DIDs need to solve a large number of candidate
models of other agents that represent how the agents optimize their decisions
in an uncertain environment. In addition, the I-DIDs track the evolution of
all the models as other agents observe, act and update their beliefs over time.
Consequently, the computational complexity of solving I-DIDs is mainly due
to the exponential growth in the number of models that are ascribed to other

c Springer International Publishing Switzerland 2015
L. Cao et al. (Eds.): ADMI 2014, LNAI 9145, pp. 1–11, 2015.
DOI: 10.1007/978-3-319-20230-3 1
2 Y. Pan et al.

agents. The complexity is further increased when a large number of agents are
to be modeled in the I-DIDs. The existing research of I-DID mainly focuses on
the case of n = 2 agents, which is not a general setting in practical applications.
In this paper, we extend I-DIDs for solving interactive decision making prob-
lems with n > 2 agents. Following the conventional representation of I-DIDs,
we need to introduce additional model space to represent every of other agents.
This not only increases the modeling complexity, but also grows the solution
complexity due to the increasing number of other agents’ models. Observing pos-
sible relations of agents’ actions, we proceed to reduce the modeling complexity
by simplifying the extended I-DID representation. We construct their relations
using Bayesian networks [8] and learn the model parameters accordingly.
We organize this paper as follows. We briefly review the I-DID model
in Sect. 2. Subsequently, we extend I-DID for the case of multiple agents
(n > 2) in Sect. 3. One learning algorithm is proposed to simplify the extended
I-DID in Sect. 4. We show the preliminary results by experimenting the proposed
techniques in one problem domain in Sect. 5. Additionally, we review the rele-
vant work in Sect. 6. Finally we conclude the paper with remarks on the future
work.

2 Background: Interactive DID

We briefly describe I-DID as well as the previous algorithms for solving the
model, and refer the reader to [4,16] for more details.

2.1 Representation
In Fig. 1, the I-DID represents how a subject agent i optimizes its decisions while
interacting with another agent j whose actions may impact their common states
S. Different from regular chance, decision and utility nodes in DID [13], the
new type of node called the model node, Mj,l−1 , models how another agent j
makes its decisions simultaneously in level l − 1. More explicitly, it contains all
possible j’s models whose solutions generate the predicted behavior Aj , which
is represented by a policy link (the dashed line) connecting Mj,l−1 and Aj . Each
model, mj,l−1 , could be either a level l − 1 I-DID or a DID at level 0 where agent
j does not further model agent i.
As agent j acts and receives observations over time, its models are updated to
t
reflect their changed beliefs. The model update link, a dotted arrow from Mj,l−1
t+1
to Mj,l−1 in Fig. 1, represents the update of j’s models over time. We zoom in
the model update link in Fig. 2. The updated models differ in the initial beliefs
that are computed through a pair of j’s actions and observations. Consequently,
the set of updated models at time t + 1 will have up to |Mtj,l−1 ||Aj ||Ωj | models.
Here, |Mtj,l−1 | is the number of models at time step t, |Aj | and |Ωj | are the
largest spaces of actions and observations respectively.
I-DID becomes a regular DID when the model update link is replaced with
dependency links and chance nodes. We may employ any DID technique to solve
an I-DID.
 Learning Agents’ Relations in Interactive Multiagent Dynamic Influence 3

Ri Ri

Ait Ait+1
t
Aj Ajt+1

t t+1
S S

Mj,l-1t Mj,l-1t+1

O it Oit+1

Fig. 1. A generic two time-slice level l I-DID for agent i.

Ajt
Mj,l-1t Mj,l-1t+1 Ajt+1
St
St+1
Mod[Mjt+1]
t
Mod[Mj ]
Ojt+1
Ait mj,l-1t+1,1
Aj1
mj,l-1t,1 1
Aj Oj1 mj,l-1t+1,2
Aj2
mj,l-1t,2 mj,l-1t+1,3
Aj2 Oj 2 Aj3
mj,l-1t+1,4
Aj4

Fig. 2. Implementation of the model update link e.g. two models, mt,1 t,2
j,l−1 and mj,l−1 ,
are updated into four models (shown in bold) at time t + 1.

2.2 Solutions

As indicated in the modeling phase above, solving a level l I-DID requires the
expansion and solution of j’s models at level l−1. We outline the I-DID algorithm
in Fig. 3. Lines 4–5 solves j’s models for the policy link while lines 7–15 imple-
ments the model update link in I-DID. Finally, lines 17–18 solve the transformed
I-DID through standard DID algorithms.
The difficulty arises for solving I-DID of a large planing horizon, T , since a
large number of j models need to be accommodated and resolved in the model
node. A set of successful algorithms have been proposed [16], and mainly focus
on prune the models of Behavioral Equivalence whose behavioral predictions for
agent j are identical [11]. Line 6 presents such a technique PruneBehavioralEq
(Mj,l−1 ) and returns representative models of j, which generates a series of
efficient I-DID algorithms.

3 I-DIDs for n > 2 Agents

We generalize the I-DID representation for the case of n > 2 Agents. Following
the notations in Fig. 4, we depict a general two time-slice I-DID in Fig. 4 where
the red block model another one agent k except agent j.
4 Y. Pan et al.

Fig. 3. Algorithm for exactly solving a level l > 1 I-DID or level 0 DID expanded over
T time steps.

Similarly to solve I-DIDs for n = 2 agents, we need to solve all candidate

models of agent K and update them in Fig. 5. This obviously increases the
solution complexity due to the growth of state numbers within additional Mk
nodes in the I-DIDs.
One natural technique is to remove the model space introduced by additional
agents. We observe that agents may act according to some certain relations.
For example, in the follow-the-leader organization, one agent intends to mimic
another agent when it interacts with a subject agent. Hence we may predict
actions of one agent according to what another agent act. Assume that there are
relations between agent k’ and agent j’s actions, we may remove Mk nodes of
agent k while relating their action nodes. We describe such reduction in Fig. 5.
In Fig. 5, the arc from Aj to Ak models the relations of actions of two agents, j
and k, and the conditional probability distributions (CPDs) model the relation
strength. In other words, agent k’s actions depends on j’s behavior. We can
predict what agent k will act according to j’s actions. The remaining problem
is to construct the relations of their actions.
 Learning Agents’ Relations in Interactive Multiagent Dynamic Influence 5

Rit Rit+1

Ait Ait+1

St St+1
t
Aj
Oit Oit+1 Ajt+1

Mod[Mjt] Mod[Mjt+1]
Oj Akt+1
Akt
Mod[Mkt] Mod[Mkt+1]

Ok

Fig. 4. Level l agent i I-DID modeling other agents j and k. The blue and red parts
represent the update of agent j’s and agent k’s models respectively.

Rit Rit+1

Ait Ait+1

St+1
St
t
Aj
Oit Oit+1 Ajt+1

Mod[Mjt] Mod[Mjt+1]

Akt Oj Akt+1

Fig. 5. Level l I-DID when relating agent j’s actions (Aj ) to agent k’s actions (Ak ).

4 Relation Learning
We consider relations of agents’ actions in a simple Bayesian network that models
influence between variables. Under this context, we aim to learn CPDs of the
arc connecting two nodes Aj and Ak . Given the known models of agents j and
k, we may solve the models to obtain their actions in each time step. Resorting
to the parameter learning techniques, we can construct the CPDs accordingly.
The learning algorithm is described in the following two steps.

Step 1: Retrieving Actions. Let Mj and Mk be two candidate sets of agents

j and k’s models respectively. We can solve the models using the standard look-
ahead techniques and represent the solutions through policy trees. A policy tree
describes the optimal decisions given observations for the entire planning hori-
zons. In Fig. 6, we plot two policy trees of time steps T =3 for two models of
agents j and k. For example, in the policy tree of agent j, the optimal decision
is a1 at the first time step t=1. At t=2, agent j takes a1 when it receives the
observation o1 ; otherwise, it takes a2 .
6 Y. Pan et al.

Fig. 6. Policy representing solutions to agents j and k’s models.

Given the policy trees of agents j and k, we can aggregate their actions
at each time step, which is indicated by the red block in Fig. 6. Subsequently,
we need to construct the relations of their actions in each time step.
Step 2: Learning CPDs. Assume that agent k’s actions depend on agent
j’s actions, we may construct Bayesian network (BN) relating their actions in
Fig. 7. Relations of agent j’s actions over time follow the update of agent j’s
models in I-DID while agent k’s actions are predicted according to j’s actions.
Consequently, we don’t need to represent the update of agent k’s models over
time. Instead we learn the CPDs, like P r(Atk |Atj ), in the constructed BN.
We follow the maximum likelihood estimation techniques to learn the CPDs
in the BN. The CPDs represent the relation strength between agents j and k’s
actions. They provide the input into the I-DID construction for n > 2. Using
the example of agents j and k, we summarize the two steps in Fig. 8.
The above procedures may replace lines 3–5 in Fig. 3 so that I-DID can
be extended to represent interactive decision making for more agents. Given
W other agents, the number of model nodes will be W in the I-DID and each
model node contains up to |M0j |(|Aj ||Ωj |)t models. Through learning relations of
agents’ actions, we may maintain only one model node in the I-DID while relating
actions of other agents. This significantly reduces the solution complexity.

5 Experimental Results

We implemented the procedures in Fig. 7 and built the I-DID for three-agent in
the Tiger problem (|S| = 2, |Ai | = |Aj | = |Ak | = 3, |Ωi |=6, |Ωj | = |Ωk |=3) [9].
We compare four methods for solving the I-DIDs.

– NO: we solve the I-DIDs exactly without reducing the model space;
– DMU: we solve the I-DIDs using the PruneBehavioralEq like the approxi-
mate DMU method [3];
– PL: we use the BN learning techniques (in Fig. 7) to relate agents’ actions
thereby reducing the model space;
– DL: we combine DMU and PL to solve the I-DID.
 Learning Agents’ Relations in Interactive Multiagent Dynamic Influence 7

Fig. 7. A Bayesian network (red parts) relates agents j and k’s actions over time.

Fig. 8. Procedures for solving agents j and k’s models and learning the relations of
their actions.

We ran experiments on a Linux platform with Intel Core 2.2.4 GHz with 4 GB of
memory. Table 1 shows the running time for solving the I-DIDs through different
methods. The time required by solving and expanding three-agent I-DIDs is
much more than that needed for solving two-agent I-DIDs. Both DMU and PL
methods significantly reduce the times particularly in complex problems like
T = 5 cases with large N . Combining two methods results in more efficient
solutions and is scalable to solve more complex problems.
We further show the running time for learning BN in the PL method. Table 2
shows that the BN learning procedure spends relatively short time and the learn-
ing time does not increase significantly with the initial number of models (N )
as well as the planning horizons (T ).
In Fig. 9, we show the average rewards gather by executing the policy trees
obtained from solving the level 1 I-DIDs for the multiagent tiger problem. We let
the number of the initial models labeled be either 25 or 50. Each data point is
the average of 1000 runs of executing the policies. In Fig. 9, as can be seen that
the average of N = 50 is larger than one of N = 25, and the average rewards
for T = 5 is larger than that of T = 3. In other words, the rewards improve
as either the number of the initial models or the planning horizon increases.
8 Y. Pan et al.

Table 1. The running times for solving I-DID for more agents.

Model Agent T Time(s)

number (N) number (W)
NO DMU PL DL
Solve level 0 models 25 2 3 63 94 * *
50 2 3 125 172 * *
25 2 5 422 532 * *
50 2 5 829 1047 * *
25 3 3 141 141 125 172
50 3 3 250 250 250 344
25 3 5 1047 890 906 1125
50 3 5 1781 1688 1735 2162
Build level 1 model 25 2 3 359 15 * *
50 2 3 1641 15 * *
25 2 5 18125 31 * *
50 2 5 139640 31 * *
25 3 3 812 110 422 16
50 3 3 3828 110 1688 16
25 3 5 71594 8019 22141 47
50 3 5 * 8047 * 47

When N = 50 and T = 5, we can’t obtain the solution using the No and PL

approaches. The PL method even achieves better rewards. By relating agents’
actions, we may focus on more influential actions of other agents.

6 Related Works

I-DIDs generalize the influence diagrams to multiagent settings facilitating deci-

sion making in the presence of other sophisticated decision makers of uncer-
tain types. They may be viewed as graphical counterparts of finitely-nested
interactive partially observable Markov decision processes (POMDP) [6]. Other

Table 2. The running times used to learn relations of agents’ actions.

Model number (N) T Time(s)

25 3 62
25 5 78
50 3 78
50 5 93
 Learning Agents’ Relations in Interactive Multiagent Dynamic Influence 9

Fig. 9. The average rewards agent i gathers when it plays with both agents j and k
according to the I-DID solutions.

prominent generalizations of IDs to multiagent environments include multiagent

influence diagrams (MAIDs) [10] and networks of influence diagrams (NIDs) [5].
Doshi et al. [4] observe that both MAIDs and NIDs provide an analysis of the
game from an external viewpoint, and adopt Nash equilibrium as the solution
concept.
As we mentioned before, a predominant factor in the complexity of I-DIDs
is due to the exponential growth in the candidate models over time. Using the
heuristic that models whose beliefs are spatially close are likely to be BE, Zeng
et al. [4,17] employed a k-means approach to cluster models together and select
K representative models in the model node at each time step. However, the
K representatives are not guaranteed to be behaviorally distinct due to which
the set is not minimal. By discriminating between model updates, the approach
generates a minimal set of models in each non-initial model node [3,16]. Another
attempt on approximating BE is to identify the BE by comparing probability
distributions over the subject agent’s action-observation paths, which however
turns out to be internally contradictory [1,2].
In a different approach, Zeng et al. [14] sought to cluster models by compar-
ing the K-most probable paths in the policy trees. However, computing path
probabilities becomes computationally hard as the paths become longer, and
bounding the prediction error is not possible. We may further prune the model
space by clustering models whose predicted actions at a particular time step are
identical [15,16]. Various I-DID solution techniques that exploit the notion of
equivalence were recently compared [16] and their effectiveness demonstrated on
several problem domains including in GaTAC. More recent research on devel-
10 Y. Pan et al.

oping partial policy trees for speeding up the I-DIDs and investigating online
planning techniques can be found in [18–20], respectively.

7 Conclusion

It is a new challenge discussed by many researchers that the I-DIDs are gen-
eralized to the setting of n > 2 agents. Instead of literally developing models
for every other agent, we resort to the BN learning for relating actions of other
agents. The proposed method simplifies the I-DID representation and signif-
icantly reduces the solution complexity. Experiments in one problem domain
show promising results. Further work would focus on more comprehensive tests
in larger scale problem domains.

Acknowledgment. This research was supported by the Nature Science Foundation

of Jiangxi Province, China (No. 20132BAB211026), and the Research Foundation of
Education Bureau of Jiangxi Province, China (No. GJJ12741 and No. GJJ13306).
The Project was also supported by the National Natural Science Foundation of China
(No. 61375070 and No. 61402306).

References
1. Chandrasekaran, M., Doshi, P., Zeng, Y.: Approximate solutions of interactive
dynamic influence diagrams using -behavioral equivalence. In: International Sym-
posium on Artificial Intelligence and Mathematics (ISAIM) (2010)
2. Doshi, P., Chandrasekaran, M., Zeng, Y.: Epsilon-subject equivalence of models
for interactive dynamic influence diagrams. In: WIC/ACM/IEEE Conference on
Web Intelligence and Intelligent Agent Technology (WI-IAT) (2010)
3. Doshi, P., Zeng, Y.: Improved approximation of interactive dynamic influence dia-
grams using discriminative model updates. In: AAMAS, pp. 907–914 (2009)
4. Doshi, P., Zeng, Y., Chen, Q.: Graphical models for interactive pomdps: represen-
tations and solutions. J. Autonom. Agents Multi-Agent Syst. (JAAMAS) 18(3),
376–416 (2009)
5. Gal, K., Pfeffer, A.: Networks of influence diagrams: a formalism for representing
agents’ beliefs and decision-making processes. J. Artif. Intell. Res. 33, 109–147
(2008)
6. Gmytrasiewicz, P., Doshi, P.: A framework for sequential planning in multiagent
settings. J. Artif. Intell. Res. (JAIR) 24, 49–79 (2005)
7. Howard, R.A., Matheson, J.E.: Influence diagrams. In: Readings on the Principles
and Applications of Decision Analysis, pp. 721–762 (1984)
8. Jensen, F.V.: Bayesian Networks and Decision Graphs. Information Science and
Statistics. Springer, New York (2001)
9. Kaelbling, L., Littman, M., Cassandra, A.: Planning and acting in partially observ-
able stochastic domains. Artif. Intell. J. 101, 99–134 (1998)
10. Koller, D., Milch, B.: Multi-agent influence diagrams for representing and solving
games. In: International Joint Conference on Artificial Intelligence (IJCAI), pp.
1027–1034 (2001)
 Learning Agents’ Relations in Interactive Multiagent Dynamic Influence 11

11. Pynadath, D., Marsella, S.: Minimal mental models. In: Twenty-Second Conference
on Artificial Intelligence (AAAI), Vancouver, Canada, pp. 1038–1044 (2007)
12. Seuken, S., Zilberstein, S.: Formal models and algorithms for decentralized deci-
sion making under uncertainty. J. Autonom. Agents Multi-Agent Syst. (JAAMAS)
17(2), 190–250 (2008)
13. Tatman, J.A., Shachter, R.D.: Dynamic programming and influence diagrams.
IEEE Trans. Syst. Man Cybern. 20(2), 365–379 (1990)
14. Zeng, Y., Chen, Y., Doshi, P.: Approximating behavioral equivalence of mod-
els using top-k policy paths (extended abstract). In: International Conference on
Autonomous Agents and Multi-Agent Systems (AAMAS), pp. 1229–1230 (2011)
15. Zeng, Y., Doshi, P.: Speeding up exact solutions of interactive influence diagrams
using action equivalence. In: International Joint Conference on Artificial Intelli-
gence (IJCAI) (2009)
16. Zeng, Y., Doshi, P.: Exploiting model equivalences for solving interactive dynamic
influence diagrams. J. Artif. Intell. Res. (JAIR) 43, 211–255 (2012)
17. Zeng, Y., Doshi, P., Chen, Q.: Approximate solutions of interactive dynamic influ-
ence diagrams using model clustering. In: Twenty Second Conference on Artificial
Intelligence (AAAI), Vancouver, Canada, pp. 782–787 (2007)
18. Zeng, Y., Doshi, P., Pan, Y., Mao, H., Chandrasekaran, M., Luo, J.: Utilizing
partial policies for identifying equivalence of behavioral models. In: Twenty-Fifth
Conference on Artificial Intelligence (AAAI), pp. 1083–1088 (2011)
19. Zeng, Y., Mao, H., Pan, Y., Luo, J.: Improved use of partial policies for identi-
fying behavioral equivalence. In: Autonomous Agents and Multi-Agent Systems
Conference (AAMAS), pp. 1015–1022 (2012)
20. Chen, Y., Doshi, P., Zeng, Y.: Iterative online planning in multiagent settings with
limited model spaces and PAC guarantees. In: Autonomous Agents and Multi-
Agent Systems Conference (AAMAS) (2015)
 Agent-Based Customer Profile Learning in 3G
Recommender Systems: Ontology-Driven
Multi-source Cross-Domain Case

Vladimir Gorodetsky(B) , Vladimir Samoylov, and Olga Tushkanova

St. Petersburg Institute for Informatics and Automation,

Russian Academy of Science, 39, 14-th Liniya, 199178 St. Petersburg, Russia
{gor,samovl,tushkanova}@iias.spb.su

Abstract. Advanced recommender systems of the third generation (3G)

emphasize employment of semantically clear models of customer cross-
domain profile learned using all available data sources. The paper focuses
on conceptual level of ontology-based formal model of the customer pro-
file built in actionable form. Learning of cross-domain customer profile as
well as its use in recommendation scenario requires solving a number of
novel problems, e.g. information fusion and data source privacy preserva-
tion, among others. The paper proposes an ontology-driven personalized
customer profile model and outlines an agent-based architecture sup-
porting implementation of interaction-intensive agent collaboration in
two variants of target decision making procedure that are content-based
and collaborative filtering both exploiting semantic similarity measures.

Keywords: Recommending systems · Ontology-based customer profile ·

Multiple data sources · Learning of customer profile · Agent-based archi-
tecture · Semantic similarity measure

1 Introduction
Advanced Recommendation Systems (RS) qualified as RS of the third genera-
tion (3G) emphasize employment of semantically clear model of customer cross-
domain profile learned using all available data sources where the customer’s
“footprints” can provide, for learner, with useful information about customer’s
interest and preferences. Focus on semantic aspects of customer profile stimu-
lates, in turn, wide spread of ontology-based meta-modeling of data sources. It is
worth to note that well known fact that customers prefer to trust much more to
the recommendations of their “friends” than to anonymous sources like routine
advertisements is an additional argument in favor of importance of semantics in
customer profiling. Indeed, the core of the customers’ trust to the “friends” is
their semantic similarity. As a result, e.g., collaborative filtering (CF) as applied
to the “friend” community leads to good results due to implicit meeting the
members of this community to the semantic similarity requirement.
Recent understanding of the topmost importance of the semantic basis of
customers’ motivations determining his/her preferences in buying of those or

c Springer International Publishing Switzerland 2015
L. Cao et al. (Eds.): ADMI 2014, LNAI 9145, pp. 12–25, 2015.
DOI: 10.1007/978-3-319-20230-3 2
 Agent-Based Customer Profile Learning in 3G Recommender Systems 13

these product items results in noticeable shift of the RS-related research to the
causal analysis of the customer interests and preferences, in particular, to active
research on ontology-based model of particular customer interests and model of
customer profile as a whole. The core of this shift is focus on well semantically-
grounded personalized customer interests. Let us mote that similarity measures
constituting the basis of any former versions of CF are understood as purely sta-
tistical properties. Statistical similarity measures are independent of the causes
motivating the customer’s selections. In contrast, 3G RS similarity measures
should, first of all, to explain semantically why two customers are similar or dis-
similar, although they may select the same product item. E.g. one customer can
select a movie due to its favor director, whereas another one may to do the same
choice due to the movie genre and/or leading actor team. Former CF ignores such
facts. Therefore, customer interests presented as whole customer profile have to
be clearly semantically interpretable. Let us note that such profiles should be
learnt from all available data sources.
Accordingly, several novel problems appear in modeling of 3G RS. These
problems formulated below as the questions are the followings:

– What can be an appropriate customer interest formal model covering its mul-
tiple interests in several domains?
– How this model interacts with the multiple domain ontologies peculiar to the
applications having multiple learning data sources?
– How customer interest formal model interacts with reasoning on
recommendation-related decisions?
– How semantic similarity measures of a pair of customers can look like and
how these measures interact with the formal model of customer profile?

Many other novel questions exist too, but they cannot be answered in a single
paper. This paper focuses on conceptual aspects of formal modeling of RS com-
ponents associated with the aforementioned questions while emphasizing impor-
tant role of customer profile formal model as a core of the whole 3G RS model.
Another paper topic is about the novel roles of agents supporting interactions of
RS components in the customer profile learning and decision making use cases.
Taking the agent mining approach [4], this work combines agent technology
with ontology, customer profiling and recommender systems. To make the paper
ideas and contribution more understandable, it starts with presenting of a case
study data set comprising several data sets of cross-domain nature (Sect. 2).
Afterwards, in Sect. 3, the proposed formal model of the customer profile sat-
isfying the requirements to its semantically clear interpretation is described.
This section sketches interaction of ontologies of multiple data sources with the
customer profile formal model too. Section 4 outlines the agent-based architec-
ture of RS components implementing two its basic use cases that are (1) the
customer profile learning and (2) recommendation related decision making use
cases. Section 5 provides for related work survey with the focus on the exist-
ing ontology-based customer profile models. Conclusion describes the current
progress in development of the presented components and sketches future efforts.
14 V. Gorodetsky et al.

2 Data Set Meta-model

To demonstrate the basic paper ideas in more understandable mode, Amazon
data set [1] is selected as a case study. This data set explicitly uses catego-
rization of the domain concepts that makes it simpler to model it in terms of
ontology that is widely used as a basic approach to emphasize specification of
data semantics. This data set comprises several data source, deals with sev-
eral domains and implicitly contains information needed to enable employment
of semantic-based customer’s similarity. The latter makes it possible to enrich
recommendation-related decision making with additional knowledge improving
the recommendation quality.
The meta-model of this data set represented in terms of class diagram is
depicted in Fig. 1. Its basic concept is Product specified in terms of id, ASIN
(Amazon Standard Identification Number), title, group of Products it belongs
to (Book, DVD, Video VHS or Music - see Fig. 2), salesrank (rate of Product
sales), similarid (set of other Product ASINs bought together with the Product),
categories. The categories the Product are given as a hierarchy of sequential nodes
separated by symbol “|”. Category Id is indicated within the squired brackets
[*]. Product is also assigned with customers reviews, which attributes are time,
customer’s Id, rating value, total number of reviews. The last attribute of the
concept Product is avgrating that is averaged Product rating assigned to it by
customers, in their reviews.

Fig. 1. Amazon data set class diagram structure

Figure 2 presents additional information about the Amazon data set structure
while depicting entity-relationship diagram with extended information about the
following concepts: Product group categories, Product similarity and customer’s
 Agent-Based Customer Profile Learning in 3G Recommender Systems 15

Fig. 2. Entity-relationship class diagram of the Amazon data set

reviews that can be provided for a Product. Figure 3 gives a shortened example
of the Amazon data set record specifying an instance of the Product of the
group Book and representing the instance properties related to categorization
and similarity.

3 Structured Representation of Customer Profile

Ontology categorization presenting concept hierarchy is the first class component
of any ontology. Let us note that each Product instance can belong to several
categories (see, for example, Fig. 3).
What is important, that categorization of the ontology concepts can be intro-
duced in many different ways. For example, some, perhaps, artificially introduced
subcategories can be of great importance for practical purposes. One of such
examples can be seen in Amazon data set ontology fragment shown in Fig. 4.
Each node of this fragment represents a subclass of movies. It is worth to note
that each node of the ontology hierarchy can be uniquely mapped a subset of
Product items (a subset of particular Movies, in Fig. 4) possessing corresponding
properties (in Fig. 4 these properties shown inside the blocks).
Vice versa, let us assume that the items of Product are mapped to the cat-
egories introduced in ontology (like the one done in Fig. 4) correspond to such
Product items that were selected and positively reviewed by the target cus-
tomer. In such cases, every particular node will represent the favor customer
selections that can be used as learning data set to discover the properties of
the items selected by the customer and some of the discovered properties can
16 V. Gorodetsky et al.

correspond to a particular customer interest. The node corresponding to the

favorite customer selections with explicitly indicated common properties discov-
ered via learning can be interpreted as a particular category of the ontology.
Ontology comprising such node-categories can be considered as a concept hier-
archy presenting customer interests in structured form. Such structure would be
considered as a useful variant of the customer profile. The question is how to
select such subcategories of Product that fit the customer interest structure in
the best way.

Fig. 3. An example of the Amazon data set record specifying an instance of the Product
of the group Book concerning with categorization and similarity

Fig. 4. Customer-profile-oriented ontology hierarchy for concept Movie

Fortunately, this task is not new. Let us remind that the routine subtask of
Machine learning that is selection of informative features if successfully solved
results in a set of features and a quality measure of each such feature is well
 Agent-Based Customer Profile Learning in 3G Recommender Systems 17

known measure called coverage. At that, usual requirement to the found set of
the features is that the found features together have to provide for coverage of all
learning data instances. A peculiarity of the learning task formulated in previous
paragraph, in comparison with the general case, is that the features in question
have to be presented in a special form: they have to be specified in terms of
predicates PS (xi1 ∈ X i , . . . , xi ∈ Xi ), where xi , . . . , xi – particular proper-
1 k k 1 k

ties of the concept Product, Xi1 , . . . , X i – sub-domains of the above properties
k

domains, and predicates PS (xi1 ∈ X i , . . . , xi ∈ X i ) take the value true if and

1 k k
only if the membership indicated in the parenthesis is held.
Thus, the conclusion resulting from the above text is that if the customer
interests are expressed in terms of a set of predicates PS (xi1 ∈ X i , . . . , xi ∈
1 k

Xi ) then the customer interests are specified in term of structured specific ontol-
k
ogy sub-categories. As a result, they are clearly interpretable in terms of ontol-
ogy concepts and their attributes are given as some statements about Product
properties. In this case, the remaining problem is whether a Machine learning
technique that is capable to discover knowledge representing formally customer
interests in terms of predicates PS (xi1 ∈ X i , . . . , xi ∈ Xi ) exists. The answer
1 k k
on this question is positive: a variant of such a Machine learning technique was
proposed in [6]. Since the description of this technique is out of the paper scope,
it is omitted here and the interested are referred to [6].
Hereinafter, it is assumed that the customers profile is represented in the
form of a structured subset of domain ontology concepts and looks like it is
shown in the toy example depicted in Fig. 4 with each node NS associated a
set of Product instances that match the properties indicated by the predicate
PS (xi1 ∈ X i , . . . , xi ∈ X i ) mapped to the corresponding node NS . Figure 5
1 k k
extends Fig. 4 in such way while exemplifying the structure of the customers
profile model considered in this paper.
Several advantages are peculiar to the proposed formal model of the customer
profile. Some of them are as follows:
1. It is compact, clearly and unambiguously interpretable in semantic terms
of ontology concepts: each particular interest is a subclass (a category) of the
domain ontology.
2. This formal model naturally implements personalization.
3. Profiles of various customers in the same domain are presented in terns
of the same concept subclasses and therefore they are simply comparable since
computing a semantic similarity of a customer pair requires to compare the both
profile structures and to find the set of common successors. Semantic similarity
measure can be expressed in terms of relative number of common interests of a
pair of customers peculiar to the Product item under recommendation procedure.
In fact this measure should be a subject of special research and experiments.
4. Customer’s profile, in fact, is represented in about actionable form: only
few efforts are required to design decision making mechanism, e.g. decision tree,
voting mechanism or some other one.
18 V. Gorodetsky et al.

Fig. 5. Formal model of customer profile fragment: An example

5. Context can be represented via particular component of the domain ontol-

ogy, and therefore, the proposed customers profile model can be naturally gen-
eralized to the case of context-aware recommendations.
6. The same concerns with the cross-domain recommendation technology:
in any case, cross-domain recommendation is realistic in the cases when corre-
sponding domains have something in common that can be expressed in terms of
common concepts of both domain ontologies. Particular techniques of ontology-
based cross-domain decision making is a subject of a deeper research. For exam-
ple, Fig. 6 represents graphically a fragment of the ontology hierarchy for Product
of the group Music peculiar to a particular customer. It contains hierarchy of the
nodes entitled Comedy. Figure 7 represents a fragment of ontology group “Video
VHS ”. It also contains the nodes entitled Comedy. Comedy as a genre can be a
customer interest and both ontologies (given in Figs. 6 and 7) can contain some-
thing in common, for the same particular customer.
7. Finally, information fusion from several, perhaps, distributed data sources
is also naturally resolved if to use ontology as a meta-model defined on top of
distributed data sources that makes it possible to use any information fusion
strategy including those ones that provides for data sources with privacy pre-
serving.

4 Agent-based Architecture for Customer

Profile Learning
Two basic use cases of RS should be supported by its software compo-
nents, (1) learning of customer profile and (2) producing recommendations
when necessary. The basic source of the learning information is learning data
set, e.g. Amazon data set [1], in this paper. The objective of learning use
case is learning of the customer profile in the form outlined above. One of the
existing technologies proposed in [6] comprises three basis steps. In the first step,
Another random document with
no related content on Scribd:
75 MM. GUN CARRIAGE MODEL OF 1916.
LEFT ELEVATION.
 75 MM GUN CARRIAGE. MODEL OF 1916.
RIGHT ELEVATION.

The trail is made in two halves of box section built of bent and
riveted steel plate. Each half is bolted to a lug on the equalizing gear,
so that it may be rotated horizontally from the junction point of the
trail to the point where the trail hits the wheel.
The trails are locked together in traveling position by means of a
cone-shaped vertical lug on the lunette bracket which fits in a socket
in the trail coupling, and is locked in place by the trail-coupling latch.
Trail-coupling latch has a handle and catch with a vertical spindle
seated in a socket in the lunette bracket. A handle-return spring is
assembled around the spindle and the latch engages a catch on the
trail coupling when trails are fixed in the traveling position. Latch is
opened by moving handle forward.
Lunette consists of a ring for attaching the carriage to the limber
and is bolted through the lunette bracket.
 Floats are attached to the bottoms of both trails at their rear ends,
consisting of flanged steel plates for the purpose of increasing
bearing area of the trails on soft ground.
Spade bearings are riveted to rear of the trails and form bearings
for spades in firing position. Spades are driven through the bearings,
and their upward movement relative to the trails is prevented by
spade latch.
Spade-latch bracket consists of a bronze plate with a cylindrical
chamber for a spring and plunger and two bearings for latch-handle
pin. Bracket is riveted to the inside top of trail in front of the spade.
Spade-latch plunger, with a spring assembled around it, is seated in
the chamber and the spade-latch handle is pinned in the bearing.
Top of handle extends through the trail and is roughened for use as a
foot pedal. Lower part of handle engages with the plunger. When the
spade is driven the plunger is forced into a notch in the spade by
means of the spring, and the slope on face of plunger allows a
downward movement of the spade and prevents upward movement.
To release spade the foot pedal on latch handle is pressed down,
disengaging plunger from spade, and the spade is removed.
Trail handles are riveted to outside of both trails for lifting trails.
Name plate is riveted to outside lower left trail. It is important that the
number of carriage on this plate be recorded by the officer in charge
of the unit to which it is assigned and that this number be used as a
reference in all correspondence. Wheel guards, rear, are plates
riveted to the outside lower left of both trails for the protection of trail
bodies against contact with limber wheels on short turns. Trail
guards are bent plates riveted to the top of trail in front of trail-
coupling latch to prevent battering of trails by sledges used for
driving the spades.
Sponge-staff fastenings are riveted to tops of both trails. Sponge
staffs are inserted in upper rings of staff fastenings and the lower
ends are clamped in place. The smallest section of sponge staffs fits
in sponge fastenings.
Sledge fastenings are similar to sponge staff fastenings and are
riveted to the outside of each trail. Wheel guards (front) are plates
riveted to the outside of trails near the front to prevent contact of
trails with wheels when the trails are separated.
Spare parts case is a steel box with a hinged steel cover provided
with a bolt snap and padlock riveted to the outside of front left trail.
This case contains spare parts for emergency use.
Trail seats are made of formed bent plates riveted to the tops of
trails near breech of gun. Oiler support with springs is under the
right-hand trail seat. Oiler rests on this support and is held in place
by springs.

75-MM. GUN CARRIAGE, MODEL OF 1916.

REAR VIEW.
 75 MM. GUN CARRIAGE, MODEL OF 1916.
PLAN VIEW.

Traveling lock bar consists of a forged steel bar pinned to lock bar
bearing on left trail and made to swing across trails in traveling
position and along left trail in firing position. In traveling position the
socket in the middle of the lock bar engages with the traveling lock
stud in the bottom of cradle, and right end of lock bar is held in lock
bar clip on right trail by the latch. To disengage the latch for firing, the
latch handle is lifted and the lock bar swung to fastening in left trail,
where it latches.
To lock the cradle, the gun is brought to 0 azimuth and the
traveling lock pointer on right trunnion cap brought to line marked
“March.” In this position the traveling lock socket fits over stud, and
the lock is latched. The latch consists of a lever pinned at one end to
the lock bar with a plunger pinned in center extending through the
bar with a spring around the plunger body to hold the latch in place.
Trail connections are riveted to front end of trail and bolted to
equalizing pinions.
 The cradle comprises the spring cylinder with attached parts.
The spring cylinder is below and shorter than the gun. It is in the
form of two cylinders joined at the center, with axes in the same
horizontal plane. Above the cylinders are the gun ways, parallel to
the cylinders, bronze lined, and opening toward the center line of
cylinders. Traveling lock stud is bolted through a lug at the rear and
below the cylinders. Firing-shaft bracket is riveted to the left side and
range-scale bracket to the right side of the cylinder at its rear end.
Shoulder guards are pinned in sockets in both firing-shaft and range
scale brackets to prevent contact of the gun during recoil, with the
cannoneers. Trunnions are riveted and keyed to the cylinder near
center. Elevating arc is bolted to lugs on the bottom of cylinder at
trunnions. Piston-rod bracket is riveted to projections on the cylinder
above the gun slides near the front end. Cylinder cover is pinned to
cylinder clips, which are riveted to the front of spring cylinder. (Note:
On some carriages the clips are made integral with the cylinder.)

75 MM GUN CARRIAGE, MODEL OF 1916.

LONGITUDINAL AND TRANSVERSE SECTIONS
 The recoil mechanism is designed for variable recoil, the length
of which is regulated automatically by the elevation of the gun. The
following table gives lengths at various elevations: (These lengths
are based on theoretical calculations. Actual lengths of recoil
between 8’ and 45’ elevation are generally greater.)
Elevation. Length of Recoil.
-7.0 to plus 8.0 degrees 46 inches.
-8.0 to plus 16.47 degrees 46 to 28 inches.
-16.47 to plus 27.20 degrees 28 inches.
-27.20 to plus 36.7 degrees 28 to 18 inches.
-36.7 to plus 53 degrees 18 inches.
The breech of the gun on short recoil will strike the ground at the
level with the bottom of the wheels at an elevation of 47 degrees or
over.
The recoil mechanism is of the hydraulic spring type, with the
recoil cylinder mounted above the gun and the counter-recoil springs
in the cradle below the gun. The recoil cylinder is held in place by a
slot machined in the gun jacket at the front and rests in the
cylindrical opening in the gun lug above the rear of the gun. It is held
in place by the cylinder retainer, which screws into the rear cylinder
parallel to the center line.
The recoil valve is a cylinder with a collar at the front end and
three lands inside and parallel to the bore. Three rows of holes are
bored at the lands. The recoil valve fits inside the cylinder, resting on
the lands, and is held in place by a collar bearing against the edge of
the counterbore in the cylinder at the front, while the rear end of the
valve bears against the inside rear end of the cylinder.
The piston is screwed and pinned to the piston rod and is of
bronze, slotted to fit lands and grooves in the recoil valve. The piston
rod is hollow for almost the entire length. The front end passes
through the gland in the cylinder head and piston-rod sleeve. The
front of the recoil cylinder is closed by the front cylinder head, which
is screwed in place with a gasket. A bronze gland with four rings of
5/ inch Garlock packing prevents leakage around the piston rod.
16

The counter-recoil buffer consists of a buffer rod screwed into the

buffer nut at the rear end of the recoil cylinder, and extending
through the buffer bushing into the interior of the piston rod. The
buffer head is screwed and pinned into the front end of the buffer
rod. The buffer head is of two diameters and connected by a short
cone. The rear end is the smaller diameter and is threaded inside to
screw over the buffer rod. The coned surface contains slots leading
to a hollow chamber in front. The front end of head is faced and
provided with a central bearing for valve stem. The bearing is
supported by webs to main body of guide. Valve stem has a stop on
rear and a valve screwed to front. Valve is faced to seat on front of
the bearing, webs and circular face of main body of guide.
The counter-recoil springs are assembled around spring rods in
spring cylinder. Spring rods fit in gun lug and are fixed in place by
taper keys driven diagonally through lug and rod. The rod is hollow
for entire length, except at the rear, where the outside diameter is
decreased to permit entrance in gun lug. Collars are screwed and
pinned to front ends of rods. Three coils of inner counter-recoil
springs are assembled over the spring rod, surrounded by three coils
of outer springs. Inner and outer springs are coiled in opposite
direction to prevent nesting, and sets of coils are separated by a
bronze separator. Rear ends of cylinder are bushed for spring rods.
The operation of recoil mechanism is as follows:
When the gun is fired it moves back in slides on cradle, carrying
with it spring rods, buffer rod, recoil cylinder, and recoil valve. The
piston, piston rod, and spring cylinder remain stationary, being fixed
to carriage.
 75 MM GUN CARRIAGE, MODEL OF 1916.
LONGITUDINAL SECTION, RECOIL AND COUNTER RECOIL MECHANISM

The recoil cylinder being full of oil, this oil is forced by the piston
through holes in recoil valve in front of piston up into annular space
between valve and cylinder and into space behind and vacated by
the piston. The hydraulic resistance caused by forcing the oil through
the holes in valve absorbs most of the recoil energy of the gun, and
the remaining energy is taken up by compression of the counter-
recoil springs and friction.
When the gun reaches the end of recoil all of the recoil energy has
been absorbed and the counter-recoil springs acting against spring-
rod piston force the gun back to battery position. The purpose of the
counter-recoil buffer is to overcome the tendency for gun to return to
battery too rapidly, at the same time allowing sufficient speed of
counter recoil to permit maximum rapidity of fire. Buffer action is
necessary, as the strength of springs required to return the gun to
battery at high elevations is greater than is required at lower
elevations.
The action of counter-recoil buffer is as follows:
 As the buffer rod moves backward in piston rod the valve in buffer-
rod head is opened by the pressure of oil in back of valve and the
vacuum in front, which forces oil into buffer chamber in front of the
buffer-rod head. At full recoil the buffer chamber is full of oil and
buffer-rod head is inside the rear end of piston rod. When springs
force gun back in counter recoil, buffer rod moves forward,
compressing oil in chamber and forcing valve closed. This prevents
escape of oil through valve and forces oil to throttle between outside
surface of buffer-rod head and inside surface of piston rod, offering
resistance to spring action and thus easing the gun into battery. The
inside bore of piston rod is tapered at front end to increase
resistance and obtain desired decrease in counter-recoil velocity.
If guns fails to return to battery after a few rounds of rapid firing, it
is probably due to expansion of oil. This may be determined and
corrected by loosening filling plug. If oil spurts out, allow it to run until
gun is back in battery. It may be necessary to relieve oil two or three
times immediately after filling. Gun should never be allowed to
remain out of battery more than 1 inch on counter recoil without
determining and correcting the cause.
If gun remains out of battery and the relief of oil does not cause it
to return, it is due to:
(a) Weak or broken springs; (b) piston-rod gland too tight; (c) dirt
or lack of lubrication in gun slides; (d) distortion of gun on gun ways;
(e) distortion of piston rod due to improper counter recoil action.
The majority of cases are due to (a), (b) and (c).
(a) Can be determined only by removing springs, and should be
undertaken only after all other methods have been tried.
(b) Can be determined by loosening piston-rod gland. If gland is
too tight, gun will return to battery when it is loosened. If gland
cannot be loosened, piston-rod is probably distorted.
(c) Flood slides with oil, and if possible retract gun and examine
gun ways and slide for dirt.
(d) If possible allow gun to cool for 15 or 20 minutes. In case of
(a), (c) or (d) gun can generally be pushed back into battery by hand.
 (e) If piston rod or interior mechanism is distorted, mechanism
must be disassembled and defective parts replaced. If distortion has
occurred, it can generally be identified by very rapid counter recoil
for round on which gun does not return to battery. This may be
caused by foreign matter in oil causing buffer valve to stick, or by
lack of sufficient oil. If distortion has occurred, it will be near gland
and can generally be felt by running hand along rod from bracket to
gland.

75 MM. GUN CARRIAGE MODEL OF 1916

VALVE TURNING GEAR AND RECOIL CYLINDER ASSEMBLED.

In case of any improper functioning of recoil mechanism during

recoil or counter recoil, cease firing until cause has been determined
and corrected. A piece is out of action when recoil mechanism is not
operating properly and will almost certainly be damaged seriously if
further firing is attempted.
After dismounting any part of recoil mechanism or filling recoil
cylinder, gun is to be retracted and released to allow counter recoil if
possible. In performing this test, valve-turning mechanism must be
disconnected and valve turned to correspond to an elevation of
carriage of 53° before gun is retracted. Gun must not be held out of
battery more than 10 seconds before being released.
Variable recoil is obtained by varying the area of effective
throttling holes in the recoil valve. An arm on the trunnion cap is
connected by means of connecting rod, valve-turning arm, valve-
turning gear, and a piston-rod gear, to the piston rod itself. As the
gun is elevated the relation of the cylinder to the trunnion changes,
causing the piston rod to turn by means of the valve-turning
mechanism. Slots in the piston engage lands in the valve, causing
the valve to turn with the piston. As the cylinder remains stationary
the location of the lands inside of the cylinder change with relation to
the three rows of holes in the valve, and these rows of holes are
covered to produce variations in the length of recoil. At long recoil all
the rows are uncovered; at intermediate recoil one row is uncovered;
and at short recoil two rows are uncovered. The setting of the valve
in degrees elevation is shown by the scale on the piston-rod sleeve
and index mark on the edge of the piston-rod bracket bushing at the
top of the piston rod.
The top carriage carries trunnions of the spring cylinders and
rests on pintle bearing. The top carriage bears on the circular bronze
slides in upper part of pintle bearing and is centered on the bronze
pintle collar of the pintle bearing.
The pintle bearing carries the top carriage, the equalizing pinions
and the equalizing gear, and is supported by the axle arms, which
are shrunk in the arms of the pintle bearing. Axle arms bear in the
wheels.
The object of the equalizing gear is to increase the stability of the
carriage in firing when the wheels are at different elevations.
Equalizing gear is an H-section with bevel tooth sector on each end
and bronze-bushed bearing in the center. It bears over the vertical
journal below the pintle bearing and is held in place by equalizing-
gear support screwed inside the journal. Vertical deflection is
prevented by the equalizing-gear bolts which are fixed to the pintle
bearing by means of nut and shoulder, pass through slots in
equalizing gear, and support gear on bolt heads. Equalizing pinions
are bevel pinions sectors, bronze bushed, bearing over the arms of
the pintle bearing, and have the lugs for trail connection bolts.
Pinions are held in place by locking rings screwed over axle arms
and are free to revolve about the pintle bearing arms.
Equalizing pinions mesh with equalizing gear.
When the carriage is laid with wheels at different elevations, it is
more unstable than when wheels are level. If fired under this
condition, the force of recoil tends to overturn the carriage. The
function of the equalizing gear is to overcome this tendency. When
carriage is fired, firing stresses are transmitted to trails, and the side
on which the smaller stress is exerted tends to rise. This motion is
transmitted through equalizing pinion and equalizing gear to
equalizing pinion on other side, applying downward force on this trail
and preserving the stability of carriage.
The angle of site mechanism is designed to give the gun a
maximum depression of about 6° and a maximum elevation of 11°,
independent of the elevating mechanism. The mechanism is
operated by two handwheels, one on each side of gun.
 75 MM GUN CARRIAGE MODEL OF 1916.
DIAGRAM OF ANGLE OF SITE MECHANISM.

Handwheel on right side operates through bevel gear on

handwheel shaft and intermediate shaft, both mounted in angle of
site bracket, right, and cross shaft mounted in bronze bushings in top
carriage. Handwheel on left side operates through bevel gears on
handwheel shaft, mounted in angle of site bracket, left, and cross
shaft mounted in bronze bushings in top carriage. Bevel gears on
ends of both cross shafts mesh with bevel gear on angle of site
worm, which is mounted in bushings in top carriage and held in place
by angle of site-worm caps. This worm meshes with teeth cut in
rocker.
 Rocker is a U-shaped piece with bearings at the tops of both arms
and teeth cut in bottom of U. The bearings bear over and are free to
revolve about trunnions on cradle independent of trunnion bearing in
top of carriage. Top half of right bearing is formed by rocker arm,
right, which extends back and carries angle of site scale, pointer,
rack, and level and forms a bearing for elevating handwheel shaft.
Rear of rocker arm, right, is braced by rocker arm brace, a diagonal
hollow rod attached to rocker arm and rocker. Top half of left bearing
is formed by rocker arm, left, a diagonal arm extending upward to the
rear to form a support for sight. Movement of the angle of site
mechanism is limited in elevation by the rocker stop bolted to the
side of the rocker and in depression by a screw in the arc.
The elevating mechanism is designed to allow an elevation of
the gun of 42° independent of the angle of site mechanism. The
mechanism is operated by one handwheel on the right side of
carriage, which is turned in a clockwise direction to elevate gun.

75 MM GUN CARRIAGE, MODEL OF 1916.

ELEVATING MECHANISM.
 The elevating mechanism is operated through bevel gears on
elevating handwheel shaft mounted on a rocker arm, right, elevating
intermediate shaft inside rocker-arm brace, elevating cross shaft,
mounted in an elevating cross-shaft bearing bolted to the rocker, and
the elevating worm, which bears inside lower part of the rocker. The
elevating worm meshes with the elevating arc, which is bolted to the
bottom of the spring cylinder.
In indirect fire the angle of site in mils is laid off on the angle of site
scale with the pointer and the desired range of graduation brought
opposite the pointer by means of the elevating handwheel.
Operation of the Angle of Site and Elevating Mechanism. The
angle of site mechanism is operated by turning handwheel, the
movement of which is transmitted through the shafts and gears to
the angle of site worm meshing with the rocker. Movement of the
rocker is transmitted directly through the elevating worm, elevating
arc, and spring cylinder to the gun, and through the rocker arms to
the elevating mechanism, gun, cradle, and sights. The elevating
mechanism moves only gun and cradle through movement of
handwheel shafts, and the elevating worm inside the rocker, which
meshes with the elevating arc.
The angle of site scale is graduated in mils from 170 to 500. The
range scale is graduated in meters. The zero setting of the gun is
with O on the range scale opposite 300 on the angle of site scale
and the level bubble on the rocker arm, right, at the center of the
tube. This allows the maximum depression of 7 degrees (about 130
mils) or the maximum elevation of 11° of angle of site mechanism to
be read on the angle of site scale against the zero of the range
scale.
The sight, model of 1916, which acts as a support for the
panoramic or peep sight, is attached to the rocker arm, left.
75 MM. GUN CARRIAGE, MODEL OF 1916.
TRAVERSING MECHANISM.
 In direct fire, the axle of the bore is brought on the line of site by
operating the angle of site handwheel until the cross hairs of the
sight are on the target and the range is laid off independently by
bringing the desired range graduation opposite 300 on the angle of
site scale. Line of site may be set independent of the range, as there
are two angle of site handwheels.
Traversing Mechanism. The total traverse of the gun on the
carriage is 800 mils. The traversing handwheel is located on the left
side of the carriage and turns in a clockwise direction for left
traverse.
The traversing handwheel shaft is mounted in the angle of site
bracket, left, and the angle of site bracket cover, left. A bevel pinion
on upper end of the shaft meshes with bevel gear on traversing
shaft, which bears in angle of site bracket, cover, left and
intermediate shaft bearing bolted to top carriage. A bevel pinion at
lower end of the intermediate shaft meshes with bevel gear on end
of traversing-worm shaft, which is mounted in bearing in top
carriage. Traversing worm meshes with traversing rack which is
screwed to pintle bearing. Traversing stops are filister head screws
between end teeth of traversing racks to limit movement of worm in
rack.
The movement of handwheel is transmitted through shafts and
bevel gears to worm and rack. Rack is mounted in pintle bearing,
which remains stationary, and top carriage moves about its bearing
in center of pintle bearing and bronze-lined slides around the outside
of pintle bearing. Traversing scale is screwed to pintle bearing above
rack, and pointer is formed on traversing worm-shaft bearing.

DISMOUNTING AND ASSEMBLING CARRIAGE.

Note.—The first and most important precaution to be observed in
assembling guns and carriages is that all parts must be clean.
Where dismounting but not assembling operation is described,
assembling is approximately the reverse of dismounting.
 I. To remove recoil cylinder.
II. To disassemble recoil cylinder.
III. To assemble recoil cylinder.
IV. To dismount gun.
V. To remove counter-recoil spring.
VI. To remove breechblock.
VII. To replace piston rod, gland pkg.
VIII. To remove wheel.
IX. To remove shields.
X. To remove spring cylinder.
XI. To remove sight.
XII. To remove rocker and rocker arms.
XIII. To remove top carriage.
XIV. To remove equalizing gear and pinions.
XV. To remove brake mechanism.
XVI. To remove trails.

I. To Remove Recoil Cylinder.

1. Remove valve turning gear cover (take out four ⅜” bolts
attaching it to the piston rod bracket).
2. Remove valve turning gear, valve turning arm and connecting-
rod as a unit by removing split pin, nut, and connecting rod pin from
trunnion cap, right.
3. Remove piston rod (remove lash wire and two 3/16” split pins)
slide piston rod gear forward and remove.
4. Remove ¼” locking screw from top of piston rod bracket.
 5. Remove 3/16” cylinder retainer screw and loosen cylinder
retainer, but do not remove retainer.
6. Remove brass spring-rod plugs from rear ends of both spring
rods.
7. Screw spring compressor eye into rear of left spring rod. Make
loop in compressor and attach double sheave close to cradle. Attach
single sheave to lunette by means of loose cord of sheave rope.
8. Man pulling rope with from four to six men, retract gun not less
than 10”, and secure rope to lunette.
9. Remove cylinder retainer, slide cylinder forward until free of
groove in gun and remove cylinder. Handle carefully.
10. Allow gun to return to battery slowly by slacking off on pull
rope.

II. To Disassemble Recoil Cylinder.

Note.—The interior parts of recoil cylinder are made with great
accuracy to insure proper operation and must be handled with care
to avoid injury.
1. Remove recoil cylinder from carriage. (See I.)
2. Drain recoil cylinder by resting on blocks at front and rear,
removing both filling plugs and drain plug, and tipping rear end up to
allow all oil to flow out of drain-plug hole.
3. Unscrew buffer-rod nut from rear cylinder head, draw out buffer
rod until wrench can be applied on flats, and remove nut. Push rod
back into cylinder.
4. Remove lower split pin from gland lock, swing gland lock back
until free of notches in gland, and loosen gland with gland wrench.
Unscrew front cylinder head with special wrench. Threads may be
started by striking handle of wrench with soft hammer. Do not hold
cylinder in a vise.