Trav's UVa Pages

1. General

1.1 What are agents?

Here are some keywords:

Adaptive/Learning: user shouldn't have to program; should be flexible; might use NN [neural networks].
Anthropomorphic: (Foner); should resemble a person or other "real" entity; helps establish trust and expectation.
Autonomy: (Foner); not over-the-shoulder, not just a front end.
Communication: 2 way feedback (collaboration), agree on "contract", have to be able to "trust" the agent; discourse restricted by domain.
Expectations: (Foner); want operation to be predictable, intuitive; "anthropomorphic".
Graceful Degradation: (Foner); not all-or-nothing; give me the best results you can.
Mobile: (Franklin/Graesser); move to new machines; reproducing?
Proactive: (Franklin/Graesser); goal-oriented.
Reactive: (Franklin/Graesser) - respond to environment (again, dependent upon domain).
Temporally Continuous: (Franklin/Graesser); not just a one-time thing; continual refinement of results.
User-Customizable: requires memory, possibly learning.

Franklin and Graesser's definition:

An autonomous agent is a system situated within and a part of an environment that senses that environment and acts on it, over time, in pursuit of its own agenda and so as to effect what it senses in the future.

1.2 General references

These are the best on-line introductions to agents that I found. They should be read in order, because the latter refer to the former.

Larry Foner: talks about what are and are not agents, and describes Julia, a MUD robot.
Franklin and Graesser: discuss agent definitions, classifications, multi-agent systems, etc.
Charles Petrie: talks about web agents, and why Foner and Franklin/Graesser are wrong. He analyzes intelligence versus autonomy and talks about various types of agents, both real and "wanna-be".

1.3 So what's the answer; what's an agent?

There seems to be a lot of snobbery among the agent-interested community. The above authors frequently refute each other's definitions of agency, and call each other's programs "incredibly simpleminded" and "lacking initiative."

No one seems to agree on what an agent is.

2. Web Agents

Most of the so-called "Web Agents" are actually just front ends for automating web activity such as meta-searching, browsing for URL updates, etc.

Here are some of my notes on Web Agents. These are the "best" links I've found so far, in that they either have something to download, or are close conceptually to the type of agents in which we're interested (see IR Agents section).

NOTE: Most of these links no longer work. Sigh.

NASA's Sulla Program: http://rbse.jsc.nasa.gov/agents/sulla.html; interest profiles, goals, HTML/FTP/news/WAIS; still in development
Autonomy Web Researcher: http://www.agentware.com/modules/web.htm; bad links
IntFilter (Intelligent Message Filtering): http://www.dsv.su.se/~fk/if_Doc/IntFilter.html; closed, focuses on email/news
Web Agents - J. Purcell (MTSU): http://www.mtsu.edu/~jpurcell/Software/webagent.htm; downloadable web agents for Win32; most of these download pages and facilitate searches.; - WebFerret - simple meta-searcher (of no interest to us); - AutoWinNet - checks your email, web pages, etc.; - TeleportPro - searches/copies web sites
CSWL - "Building Custom Web Agents": http://199.2.24.2/demosite/homewr.htm; web search robot (filters domain); supposed to have a demo, but the link is broken
AgentSoft: http://www.agentsoft.com/; software for writing agents; includes demo applets (both client and server); client applets don't work (missing demo.Applet3 and demo.Applet6 classes); server applets seem to be simple meta-searchers

3. IR Agents

So how can we use agents in IR? Some ideas (inspired in part by Gerry's Library Agents concept):

3.1 For web searching

Examine users' Netscape bookmarks (in ~/.netscape/) and links pages (in ~/public_html/); characterize the "type" of pages a user tends to link to.

For example, if Miro links to a lot of sites that are written in Croatian, or a lot of sites that have links to files rather than links to pages, then we can weight his search results accordingly.

This type of preference characterization also applies to the amount of graphics, source code citations (and the actual coding languages), links-to-content ratios, amounts of bad language, etc.

3.2 For similarity determination

Have an agent "spider" around the Internet, looking for references to documents within a certain collection. If two or more documents are mentioned in close proximity, rate those documents more similarly (for clustering purposes).

This would also apply to queries; if query_1 is deemed to be similar to document on the Internet called remote_X (by deemed I mean we perform some sort of search of a query against the rest of the Internet) and remote_X mentions documents local_A and local_B, which exist in the collection, then you could improve local_A's and local_B's estimated "relevance" to query_1. In other word, you could search other collections to transitively deduce relevance.

4. Library Agents

4.1 Introduction

Gerry Mckiernan is trying to investigate what he calls "Library Agents." These are agents whose function makes his job as librarian easier.

He wants to have agents that look at the different departmental and faculty web pages and come up with research interest profiles (RIPs). Gerry's more interested in group profiles than individual profiles, so also terms these "collective user profiles."

These profiles would be used to help assist web-searching, as well as ferret out relevant information for each deparment (information that, given the department's profile, seems likely to be of interest).

For more specific info, see his letters to us (below).

4.2 Correspondence

07-JUN-1998: Gerry's initial letter, entitled "Content Analysis of Academic Departmental Homepages"
11-JUN-1998: Gerry's response to a letter I sent him; he goes into a little more detail in what he expects from a Library Agent.

5. Links

SICS Intelligent Software Agents: Introduction and categorization of agents. Some of these links don't work.
UMBC AgentWeb: This is a huge collection of agent-related links.
BotSpot: Articles and links for the most popular Internet 'bots (impersonation programs).

IR - Intelligent Agents - Notes

Page Contents: