WEB AGENTS FOR INFORMATION ACCESS

	 Chanda Dharap, Martin Freeman, and Doreen Cheng
	Philips Research Laboratories, Palo Alto, CA 94304 
 
The ability to use transportable information agents to interoperate and browse
wide-area information is the main motivation for this abstract. The key idea is
to use agents that understand underlying representation standards as well as
communicate over common substrates. Our work involves the construction of
information agents towards supporting interoperability, content sharing, and
information discovery. In the pursuit of this work we are facing issues that
have impact on the infrastructure supporting transportable agents in
heterogeneous environments. 

Our work[1] highlights the use of both agents and structured information to
improve resolution of dynamic and heterogeneous information content across
wide-area networks. An important feature of our infrastructure is that we use
Java as a substrate on top of which we implement Knowledge and Query
Manipulation Language KQML semantics for agent interoperatibility. 

The Java layer provides the ability to build compact, platform-independent and
scalable agent API's. Java's abstraction mechanism and techniques of
information-hiding provide a uniform and protected interface to information
providers. This enables our infrastructure to scale to diverse information
services with varied interfaces. In the implementation of the infrastructure we
provide abstractions for mapping agent requests into the provider's query
language of choice. Abstraction, as provided by Java, allows us the sharing of
programming interfaces without being completely aware of the implementation.
These considerations make Java an appropriate vehicle for delivering the agent
paradigm to the Internet, where a Java encoded agent could move across
heterogeneous hardware platforms.

Our infrastructure currently consists of a confederation of static agents and
transportable agents which communicate using KQML. Static agents are
gatekeepers for applications and may themselves perform additional functions
(e.g.; send mail, filter netnews). Transportable agents combine the
functionality of both descriptive and prescriptive interactions. That is,
transportable agents may contain data, programs, procedure calls and query
scripts encapsulated as KQML messages. In addition, they may contain
administrative information used by static agents. The descriptive syntax of
KQML makes it an expressive language for agent communication.

In our design, agents also make use of structured meta-information to compose
information of interest to the user. Structured meta-data effectively describes
content, thus enabling the agent to evaluate possible matches for user requests.
In a simplistic view, an agent accepts structured specifications for the
desired information, interacts with relevant information providers, and
composes information from various information sources to fit the user's
specification. Agents are capable of carrying state information and partial 
results from provider to provider in an effort to satisfy more complex
requests.  The use of structured data to represent content as well as 
structured encapsulation of agent communication (KQML), increases the
effectiveness of content discovery as well as significantly lowering network
load and limiting user-interaction.

Our experience has uncovered a few problems that may only be effectively
addressed by coordinated work of professionals in the community.  One problem
stems from the fact that information is not uniformly structured and is
presented to the user in diverse ways. Another stems from the fact that global
information changes dynamically and non-uniformly. It is difficult to keep
track of updates and additions to the information on the Web. 

Emerging work on data representation indicates a trend to move towards global
standards. Some examples of this can be seen in Harvest's SOIF, IAFA and
WHOIS++ templates. However, we believe that in the fast expanding information
space, enforcing any rigorous standard is an unrealistic solution. A better way
may be to use transportable agents to compose a unified view of information by
translating between diverse representations.

So also, there are many aspects to agent development, as reflected in the work
done by many in professional communities. Several agent paradigms have been
proposed, with diverse architectures, communication protocols and semantics. 
Using Java as a substrate to unify interagent communication could solve the
problem of diversity in wide-area information as well as diversity in agent
technology. 

[1] "Information Agents for Automated Browsing" - Chanda Dharap and Martin
Freeman, To Appear in the CIKM96: The Fifth International Conference on
Information and Knowledge Management, November 1996. URL
"ftp://ftp.cse.psu.edu/pub/chanda/papers/cikm96.ps.gz"