1

CHAPTER 1

INTRODUCTION

Applications requiring intensive problem-solving capabilities have

highlighted the need for dynamic behavior oriented systems. Though being
complex in nature, systems in all enterprises are needed to solve the problems
ranging from scientific studies, commercial solutions to academic endeavors.

‘Information’ is vital for the success of any enterprise and in general

any human endeavor. Acquisition, storage and retrieval of information are key
challenges faced by the information providers and consumers today. Access to
current scientific, engineering, and other technical information is the
foundation for acquiring global competitiveness in any chosen field. However,
access to publications is often too hard and / or too expensive for individuals
or libraries. This is proving to be the ‘Achilles heel’ of the entire scientific
publication process.

Usage of Information and Communication Technologies has paved way

for the creation of ‘Digital Libraries’. These libraries promise to enable the
right information to reach the right person or persons at the right time. They
provide the platform for life-long learning.

In 1993, when the National Science Foundation prepared the “Source

Book on Digital Libraries”, (Edward A. Fox, 1993) it was estimated that
 2

nearly 100,000 volumes or 50,000,000 pages must be converted into the

digital form. If an OCR is used to digitize these pages it would result in about
10 Terabytes of public information. The Internet provided access to 20-50
Terabytes of Public Information in 2000. The amount of information
accessible using the Internet in 2003 is estimated to be 167 Terabytes. This
rapid growth in the quantum of accessible digital information was possible
largely due to the falling prices of scanning devices, highly accurate OCRs
and storage devices. The ICT made rapid progress during this time period
making browsing feasible worldwide.

However, digital libraries brought into sharp focus several key

problems pertaining to the users. Improving the efficiency of retrieving the
desired information by any given user of the digital library has proven to be a
major challenge. It is often observed that information needs of a given user,
more often than not, tend to be unique to that user. Hence storing the
information to maximize retrieval efficiency for a given user population needs
clever classification schemes. This work reports a mathematical model that
results in improved efficiency of retrieval and improved relevance of the
retrieved information. The model is for geographically distributed digital
libraries. It can also be used for generic information repositories.

The mathematical model is implemented using a multi-agent system.

Information environments like the Internet require dynamic, autonomous,
optimized, intelligent, proactive and complex solution mechanisms for various
purposes. The web digital library is one such information environment and is
distributed, large, open and heterogeneous in nature. This increasing
complexity of web digital library is complemented by an increasing
complexity of their applications. The Multi-Agent System promises to play an
important role in developing and analyzing models to solve such complex
problems. An agent in a multi-agent system
 3

Decomposes its goals and tasks, allocates sub-goals and sub-tasks

to various other agents and synthesizes partial results and
solutions.
Represents and reasons about the actions, plans and knowledge in
order to solve the complex tasks.
Realizes intelligent processes such as problem solving, planning,
decision-making and learning in order to collectively carry out
such tasks in a coherent way.
Solves issues and contributed to research in the domain-specific
areas with certain restrictions.

The appropriate mathematical model implemented in this work

provides a coherent classification scheme and content allocation. This results
in a storage methodology that provides efficient retrieval for a set of user
patterns of access. The multi-agent based implementation factors these user
patterns of access in a distributed system of digital libraries. The
implementation can be used over any distributed digital repository of
information.

1.1 OVERVIEW OF INTELLIGENT MULTI-AGENT SYSTEM

An agent is a computer program (Michael Wooldridge 2002) that is

situated in some environment, and that is capable of autonomous action in
order to meet its design objectives. As an intelligent entity (Gerhard Weiss
1999), an agent operates flexibly and rationally in a variety of environmental
circumstances given its perceptual and effectual equipment. An agent, on the
basis of key processes such as problem solving, planning, decision-making
and learning achieves behavioral flexibility and rationality.
 4

In general there is no universally accepted definition of the term

‘Agent’ and that autonomy is central to the notion of agency. One of the
definitions of the notion of autonomy explains that without human interaction
the system is able to perform the task. That is, it has control over the internal
state as well as behavior. The second definition explains that in most domains
of reasonable complexity, an agent will not have complete control over its
environments. It will have at best partial control, in which it can influence.

One of the primary paradigms of Distributed Artificial Intelligence

(Gerhard Weiss 1999) system is multi-agent system in which several such
agents coordinate their knowledge and activities in order to solve a problem in
a complex network environment. In the distributed problem solving systems
the work of solving a particular problem is divided among a number of agents
that divide and share knowledge about the problem and the developing
solution. The traditional Artificial Intelligence (AI) concentrates on agents as
“intelligence stand alone systems” and on intelligence as a property of systems
that act in isolation. Distributed Artificial Intelligence (DAI) concentrates on
agents as “intelligent connected systems” and on intelligence as a property of
systems that interact. Where traditional AI focuses on “cognitive processes”
within individuals, DAI focuses on “social processes” in a group of
individuals. Where traditional AI considers systems having a single focus of
internal reasoning and control and requiring just minimal help from other to
act successfully, DAI considers systems in which reasoning and control is
distributed and successful activity is a joint effort. Where as traditional AI
uses psychology and behaviorism for ideas, inspiration, and metaphor, DAI
uses sociology and economics. In this way, DAI is not as much as
specialization of traditional AI, but a generalization of it.
 5

The key problem facing an agent is that of deciding which of its actions
it should perform in order to best satisfy its design objectives. Agent
architecture for decision-making process can be affected by a number of
different environmental properties. Russell and Norvig (Stuart Russell 2003)
suggest the following classification of environment properties.

Accessible Vs. Inaccessible

An accessible environment is one in which the agent can obtain
complete, accurate and up-to-date information about the environment’s state.
Most moderately complex environments are accessible. The more accessible
an environment is, the simpler it is to build an agent to operate in it.

Deterministic Vs. Non-deterministic

A deterministic environment is one in which any action has a single
guaranteed effect-there is no uncertainty about the state that will result from
performing an action. The physical world can do all intents and purposes. This
is regarded as non-deterministic. Non-deterministic environments present
greater problems for the agent designer.

Episodic Vs. Non-episodic

In an episodic environment, the performance of an agent is dependent
on a number of discrete episodes, with no preferences between the
performances of an agent in different scenarios. Non-episodic environments
are simpler from the agent developer’s perspective because the agent can
decide what action is necessary to perform based on the current episode. It
need not reason about the interactions between this and the future objects.
 6

Static Vs. Dynamic

A static environment is one that can be assumed to remain unchanged
except by the performance of actions by the agent. A dynamic environment is
one that has other processes operating on it and hence changes its ways
beyond the agent’s control. The physical world is a highly dynamic
environment.

Discrete Vs. Continuous

An environment is discrete if there are a fixed, finite number of actions
and percepts in it. Russell and Norvig give a chess game as an example of a
discrete environment and taxi driving as an example of a continuous one.

1.1.1 Characteristics of Intelligent Multi-Agent Systems

The characteristics of intelligent agents exploit the dynamic nature. The

characteristics of the intelligent agents can be classified in to two categories,
such as internal (intra) and external (inter) characteristics (Walter Brenner
1998). The internal characteristics are autonomy, reactivity, proactive ness,
goal-orientation, mobility, learning ability and adaptability. The external
characteristics are communication, cooperation, coordination and self-
proclamation.

1.1.1.1 Autonomy

The autonomous agents (Walter Brenner 1998, Dimitris N. Chorafas

1998, Hyacinth S. Nwana 1996b, Stan Franklin 1996) are capable of following
their goals independently. An agent must have control over its actions and
internal states. This important behavior differentiates agents form non-
intelligent agents. An agent does not need to have each of its steps approved
 7

by its user or other agents. It is capable of acting alone. Also the degree of
autonomy varies for each software agents. It must have a certain minimum
degree of intelligence in order to be designated as being an agent because it
acts as a virtual person to the user. The intelligence of an agent is formed from
three main components. (a) Its internal knowledge base, (b) Its reasoning
capability and (c) Adaptive behavior.

1.1.1.2 Reactiveness

Agent is able to react ( Stuart Russell 2003, Walter Brenner 1998,

Hyacinth S. Nwana 1996b, Stan Franklin 1996) to the specific action in the
environment. It is the fundamental characteristic of any agent. It does any one
of the specific actions defined in the agent. It is able to change its action
according to the environment.

1.1.1.3 Proactiveness/Goal-orientation

Proactiveness (Walter Brenner 1998, Hyacinth S. Nwana 1996b) is

above the level of reactiveness. It has a good understanding about the
environment, because it takes specific initiative under specific circumstances
of its own in order to accomplish its goal. This is also described as goal
orientation because the capability of an agent itself is to take initiative. It
requires that the agent has well defined goals or even a complex goal system.

1.1.1.4 Stationary/Mobility

Stationary agents (Walter Brenner 1998, Hyacinth S. Nwana 1996b,

Hiking and Climbing 1998, Hyacinth S. Nwana 1996, Katia Sycara 1996) fix
as itself in a single system for performing specific tasks. But, then it may
 8

possess the ability to communicate with other agents in the communication

network. Remote Procedure Call (RPC) is the basis for possible
implementation of stationary agents. Mobility is an important characteristic
that permits the agent to migrate to different nodes in the electronic network.
This mobile agent can be either a mobile script or mobile object. Mobile
scripts are explicitly transferred to the target node but the mobile objects are
capable of transferring itself to the target node. Remote programming (RP) is
the basis for implementation of mobile agents.

1.1.1.5 Learning ability / Adaptiveness

The agent has the ability to adapt (Walter Brenner 1998, Borking.J.J
1999, Hyacinth S. Nwana 1996) to the changes in the environment. The
agent’s reasoning power should have a certain amount of rationality. The
agent perceives the changes in the environment; it identifies the pattern of
functioning of the environment and tries to adapt that pattern.

1.1.1.6 Communication

In a Multi-Agent System independent agents (Gerhard Weiss 1999,

Walter Brenner 1998, Michael R. Genesereth 1994) are communicating with
each other. These communications are referred to as high-end
communications. There are a number of communication protocols and
sometimes these are referred to as Agent Communication Languages (ACL).

1.1.1.7 Cooperation/Coordination

Communication is the basis for this cooperation (Walter Brenner 1998,

Hyacinth S., Nwana 1996b, Katia Sycara 1996). In order to cooperate, it
 9

needs to posses a social ability, that is, the ability to interact with other agents.
A cooperation topology defined by Franklin in 1997 is shown in Figure 1.1. It
differentiates the top layer from independent and cooperative agents. If the
individual agents of a multi-agent system are completely independent of each
other and follow their own goals, it is described as an independent system. If
the independently acting agents also have non-correlated goals, then the
system is termed as discrete. Systems with emergent cooperation develop
when agents independent of each other follow the same goals and hence from
the outside, give the appearance of cooperation. For example, if several
agents independently of each other follow an identical goal, a non-
participating observer can have the impression of a cooperative methodology.
Cooperative systems have explicit cooperative mechanisms. The agents are
constructed in a way that they can cooperate with other agents to achieve their
goals. Such agents make intentional use of this capability. It is considered that
the cooperation can precede either communicatively or non-communicatively.
Communicative agents use communication protocols and procedures to
cooperate with other agents. This communicative cooperation takes place
indirectly using the environment. An agent observes its environment, and in
this manner, notices changes caused by other agents. Its response to these
changes, in turn, causes a response to other agents observing the environment.
This is the way an indirect cooperation occurs.

Communicative cooperation agents can be divided into deliberative and

negotiating agents. A common planning and agreement of the methodology
occurs between all agents in a deliberative agent system. In negotiation-
oriented methodologies, a number of agents are competing with each other to
solve a particular problem openly available in the environment. The
negotiations can resolve conflicts and assign tasks.
 10

Coordination is a centralized strategy in which a particular agent takes

the responsibility. It decomposes the task into subtasks based on the domain
specific knowledge and then delegates the subtasks to other task specific
agents. The task-specific or coordinating agent will then take the
responsibility for collecting data, resolving conflicts, coordinating the related
agents and finally report the result to the user.

1.1.1.8 Self-proclamation

It is a specific type of cooperation in which it is able to share its

knowledge voluntarily, if any specific tasks or information requirement arises.

1.1.1.9 Personal Characteristics

These personal characteristics (Walter Brenner 1998) are

trustworthiness, reliability and security. Security is the most important
characteristic for reliability and trustworthiness.
Multi-Agent System

Independent
Cooperative

Discrete
Emergent
Cooperation
Non- Communicative
Communicative

Deliberative Negotiating Self-proclamative

Figure 1.1 An Agent Cooperation Topology

 11

1.1.2 Intelligent Multi-Agent System Architectures

Architecture is described as a particular methodology for building

agents. It specifies how the agent can be decomposed into construction of a set
of component modules and how these modules should be made to interact.
There are a number of Agent Architectures developed by various people at
different occasions. These may be classified into four major categories such as
Logic based Agents, Reactive, Deliberative and Hybrid Agents (Stuart Russell
2003, Walter Brenner 1998). These classifications are purely based on various
components that an agent includes.

1.1.2.1 Logic Based Agents

The “traditional” approach to building artificially intelligent systems,

(known as symbolic AI) suggests that the intelligent behavior can be generated
in a system by giving that system a symbolic representation of its
environment, its desired behavior and syntactically manipulating this
representation. The decision-making is realized through logical deduction.

1.1.2.2 Deliberative Architecture

This follows the classical AI intelligent system design approach and has
assumed an explicit symbolic model of the environment and the capability of
logical reasoning as the bases for intelligent actions. The modeling of the
environment is normally performed in advance and forms the main component
of the agent's knowledge base. The actual conversion and selection of a
suitable representation language are particularly difficult. Because of the high
complexity of such representations, deliberative agents have only limited
suitability for use in dynamic environments. In addition to its internal
 12

symbolic environment model, it also has the ability to make logical decisions.
The architecture of the deliberative agents is shown in Figure 1.2. Although a
deliberative agent has access to information receivers, these are infrequently
used to extend the knowledge base. Thus the main criticism lies in its rigid
structure. It is not possible for deliberative agents to work in a dynamic
environment because of its complex nature. The scheduler, planner and
executor take a lot of time to perform their operations.

1.1.2.2 Reactive Agents

In the case of reactive agents, intelligent behavior is possible even

without explicit knowledge representation and abstract reasoning capabilities.
Reactive agents do not possess an internal symbolic model of their
environment and they can operate quickly and efficiently because they sense
and respond to any change in the environment using sensors and reactors. A
reactive agent must not necessarily have a complex structure to be able to act
within a complex environment.

Executor Scheduler Planner

Output
(actions)

Intentions
Manager Reasoner Goals

Desires

Input Information Knowledge base

(Perception) receiver (Symbolic environment model)

Figure 1.2 Architecture of a Deliberative Agent

 13

It suffices to precisely observe the environment and to recognize a

range of simple principles or dependencies. This knowledge is used to develop
task specific modules that are capable of continuously checking their
environment for the occurrence of specific situations and to initiate direct
reaction when such a situation occurs. The following Figure 1.3 shows the
architecture of reactive agents defined by Brooks in 1986.

Agent sensors record information, forward it to the task specific

competence modules and hence produce a reaction of the competence module,
which actuators transfer to the environment. The competence modules work in
a parallel mode. These competence modules are arranged in a hierarchical
manner. The modules located at the lower end of the hierarchy are responsible
for the basic, primitive tasks, whereas the higher modules reflect more
complex behavior patterns. A reactive agent cannot solve any task for which
no competence modules exist. Every competence module can operate
autonomously and higher modules incorporate a subset of the tasks of the
subordinate modules and hence the name Sub-sumption architecture.

1.1.2.3 Hybrid Architecture

Hybrid architecture is a combination of the reactive and deliberative

components of agents modeling to produce a more powerful model. Whereas
the reactive component is primarily used for the interaction with the
environment, the deliberative systems concentrate on the area of planning and
decision-making. The system architect is responsible for the weighting of the
individual components and assumes priority for the decisions. Hybrid systems
are normally designed as an hierarchical architecture with an increasing degree
of abstraction. The lower levels are formed by reactive systems and are used
for the acquisition of raw information. Deliberative components for long-term
goal determination and planning are used in the upper levels.
 14

Competence
Module

Competence
Module
Input Output
(Perception) (actions)
Competence
Module

Figure 1.3 Architecture of Reactive Agents (based on [Books 1986])

The Interrap architecture from Muller is a typical example of a hybrid

agent system, which is described as a multi-layer architecture as shown in
Figure 1.4. Where the agent's basic behavior pattern is implemented in the
lower layers, which corresponds to the reactive component, the deliberative
processes of goal formation, reasoning and planning are performed in the
higher layers. This principle is presented in three layers of the Interrap
architecture, the behavior based layer, the local planning layer and the
cooperative planning layer. The appropriate belief models describe every
situation. The architecture is based on number of design principles

The three-layer structure describes an agent using various degrees

of abstraction and complexity.
Not only the control processes, but also the agent's knowledge
base are multi-layered.
The control process is bottom-up, in that, a layer receives control
over a process only when this exceeds the capabilities of the layer
below.
Every layer uses the operation primitives of the lower layer to
achieve its objective.
 15

The interrap architecture consists of three major components:

Knowledge base, control unit and World interface. The agent uses the world
interface to maintain contact with its environment. The knowledge base
consists of three levels of belief models. The control unit also consists of three
layers that closely correspond to those of the knowledge base. The behavior-
based layer describes the reactive capability of the agent. It is used when time
critical situations demand fast responses. The local and cooperative planning
layers form the agent’s deliberative components. They are used when a
situation exceeds the capabilities of the behavior-based layer and demands
longer-term objectives of determination and planning.

Every control unit consists essentially of two modules: the situation

recognition/goal activation (SG) module and the planning/scheduling (PS)
module. The SG module performs all steps described as part of the
conceptional model. The PS module then performs the planning process, the
provision of intentions and the scheduling.

Knowledge Base Control Unit

Social Model
SG PS Cooperative Planning
Layer (CPL)

Mental
Model SG PS Local Planning Layer
(LPL)

World Model
SG PS Behavior based Layer
(BBL)

Sensors Communication Actuators World interface

Figure 1.4 Interrap Agent Architecture [Miller 1996]

 16

1.2 PROBLEMS AND ISSUES OF DIGITAL LIBRARIES

Digital Libraries (William Arms 2000) hold any information that can be
encoded as sequences of bits. Sometimes these are digitized versions of
conventional media, such as text, images, music, sound recordings,
specifications and designs and many more. As digital libraries expand, the
contents are rarely the digital equivalents of physical items and more often
items that have no equivalent such as data from scientific instruments,
computer programs, video games, and databases. The digital libraries can be
distributed in various physical locations. The Ontology based approach used in
this work factors the problems of digital library.

The design and development of these digital libraries meet a lot of

issues, such as digital library infrastructure and architecture, human-computer
interaction, information retrieval, information mining, ontology based
services, semantic web, multi-lingual access semi structured data
management, text classification, web information gathering, information
filtering, information extraction, hypertext and multimedia, security and
privacy, document generation and electronic publishing, collection
development and management, intellectual property, storage management,
provision of service, user communities, web cataloging, metadata and content,
digital preservation, digital archives and museums, knowledge management,
content management etc. Most of the problems are interdependent. Even
though the collection is confined to set a boundary for the digital library, most
of the problems exist in the Internet also. The characteristics (Katia Sycara
2001) of Multi-Agent Systems are that

Each agent has incomplete information or capabilities to solving the

problem
There is no system global control
 17

Data are decentralized

Computation is asynchronous.

These characteristics are dominant in case of digital library problems

and so the Multi-Agent System is taken as the best paradigm to find the
solution to this problem.

In this research there are four such problems taken for investigation and they
are
Dynamic optimal content allocation method for web digital
libraries.
Design and development of intelligent agent proxy for digital
libraries.
Text classification for digital libraries
User-adaptive retrieval for the collection of research literature.

1.3 NEED FOR INTELLIGENT MULTI-AGENT TO DIGITAL

LIBRARIES

Digital Libraries are modern social and virtual institutions for

information collection, preservation and dissemination, to be distributed across
the world. The web digital library is one such information environment and is
distributed, large, open and heterogeneous in nature. The design and
development of digital library requires addressing of many issues (Daniel
Andresen 1996, Dieter W. Fellner 2003, Frew J. 2000, Cezary Mazurek 1999,
Raj Reddy 1999, Leon Zhao J. 1999) and must be able to perform intelligent
human-cognitive tasks in order to make it highly sophisticated and effective.
This increasing complexity of web digital library enables us to identify the
best-fit technology to solve the problems.
 18

An agent is an intelligent entity that operates flexibly and rationally in a

variety of environmental circumstances given its perceptional and effectual
capacities. An agent on the basis of key processes such as problem solving,
planning, decision-making and learning achieves behavior flexibility and
rationality. In multi-agent systems several agents coordinate their knowledge
and activities and reason about the processes of coordination. In distributed
problem solving systems the work of solving particular problem is divided
among a number of nodes that divide and share knowledge about the problem
and developing solution. An agent in the multi-agent system is able to
communicate, cooperate and coordinate the task in the environment. The long-
term goal of multi-agent system is to develop mechanisms and methods that
enable agents to interact with humans and to understand interaction among
intelligent entities whether they are computational, human or both. The multi-
agent system will be an ultimate solution to the problems and issues of the
web digital library. It will also be more optimal to use such a technology.

1.4 CONTRIBUTION OF THIS THESIS

In this thesis, the different issues of a web digital library are identified
and multi-agent based solutions are proposed to solve those problems.

The key contribution of this thesis can be summarized as follows:

1. Various software tools used for the design and development of

intelligent agents are identified.
2. The problems of dynamic optimal content allocation (NP-Hard) in
the real time hierarchical network system as well as the Partially
Observable Markov Decision Problem are identified. Also, design
and development of a multi-agent based user access pattern
system to solve this problem is an innovative step and is exhibited
through a simple simulation experiment.
 19

3. It is identified that the design and development of domain-specific

component increases the performance of the system. This is
proved through the intelligent agent proxy design and simulation
experiments. Specifically increasing the content and user
semantics knowledge in the design of the proxy enable the system
to have a high-hit ratio is another innovative step.
4. A new multi-agent based concept patterned classification and
retrieval/ recommendation system is designed and developed and
this enables one to solve the problem of distributed text
classification and retrieval. Dynamically adapting the user by
building the dynamic user profile to give relevant
retrieval/recommendation is yet another innovative step in web
digital library.

1.5 ORGANIZATION OF THE THESIS

In these previous sections, the characteristics and nature of multi-agent

systems, architecture, problems and issues of web digital libraries and the need
for multi-agent system based solutions to digital libraries were discussed. The
rest of the thesis is organized as follows.

In Chapter 2, a summary of agent characteristics and nature, agent

paradigms and software tools used for development and deployment of such
systems is provided. A unique methodology for the literature survey is
chosen. In this thesis we have addressed four different problems of web digital
libraries and therefore a simple review of all these problems and issues are
explained.

In Chapter 3, a multi-agent system based user-access patterned optimal

content allocation method for web digital libraries is presented. Unique
distributed algorithms to compute the optimal cost taking into account the
 20

bandwidth and storage and a system design and sample experimental results
are also presented.

In Chapter 4, an intelligent-agent based user-access patterned proxy

design method for digital libraries is presented. Unique alternative approach
for application proxy design method specifically for digital libraries is also
presented. In this design and experiment it is clearly proved that bringing the
domain knowledge into the rational system component would increase the
performance of the system.

In Chapter 5, a multi-agent based domain-specific self-proclamative

hierarchical concept patterned text classification method for digital libraries is
presented. This unique concept matrix representation allows one to represent
the document as a collection of technical phrases. Also, a concept relativity
analysis permits us to classify the documents into appropriate hierarchy. This
development and experiment has exhibited a comparative performance with
the human judged document classification.

In Chapter 6, a user adaptive multi-agent based self-proclamative

domain-specific hierarchical concept patterned text retrieval method for digital
libraries is presented. This unique concept matrix representation allows one to
represent the user interest as a collection of technical phrases. Also, concept
relativity analysis and user access pattern learning permits one to retrieve the
documents from the system as per the user requirements adaptively. In this
design and experiment it has exhibited a comparative performance in the user
satisfaction level.

In Chapter 7, this thesis is concluded by summarizing the contribution

of the thesis and providing the avenues for further research.