Information Agents

1
Information Agents
2
References

• Agent Sourcebook, A. Caglayan, C. Harrison, Wiley
  1997
• The Internet, Intranets, and the AI Renaissance,
  Daniel OLeary, IEEE Computer, Jan 1997, pp.71-78
• Jasper Communicating Information Agents for WWW,
  J. Davies, R. Weeks, M. Revett,
  http//www.labs.bt.com/projects/knowledge/jaspaper
  .html
• AI on the WWW Suppy and Demand Agents, C. Brown,
  L. Gasser, D. OLeary, A. Sangster, IEEE Expert,
  August 1995, 10(4), pp.50-55
• IA Group in TCD review http//www.cs.tcd.ie/resear
  ch_groups/aig/iag/toplevel2.html

3
Introduction

• Information agents deal in information
• search, filter, organise, index, off-line
  delivery, notification of changes, specific tasks
• Internet agents are information agents which
  operate on internet servers.

4
Study notes

• For each example of an agent, consider
• the extent to which it exhibits or
• how such an agent could exhbit
• the features
• intelligence
• autonomy
• learning
• cooperation
• mobility
• ...

5
Why we need internet agents (Caglayan)

• Huge volume of internet information
• NB The internet is more than the Web
• Huge variety in type of information
• newsgroups, PR material, on-line databases,
  academic articles, personal Web pages,
• Quality of information varies widely
• Users get lost following hyperlinks
• Note that search engines are agents. Without them
  users can only surf - follow hyperlinks.

6
Search Agents

• These are the Web robots which continuously seek
  out and index URLs on the web.
• The search interface is not an agent - no
  autonomy, direct manipulation.

Query
Query Server
Web Browser
Response
Index Database
User
Web robot
web
Search Engine
(from Caglayan)
7
Search agent web traversal

• Breadth-first or depth-first traversal ?
• Lycos records all the urls in pages it retrieves
  and selects the next one at random, except that
  it favours
• shorter urls (parent directories) and
• urls that are referenced more often.

8
Search agents indexing

• What is indexed depends on the agent
• title, headings, keywords, first few lines of
  text, any text in META tags, all text, ...
• Indexing technique
• typically all words in the indexed portions
• relatively crude indexing, compared with
  filtering agents.

9
Filtering agents

• Provide information that matches a users
  long-lived interest profile.
• Typically use a limited number of possibly
  heterogeneous sources
• but index the information fully (IR techniques)
  for more precise retrieval.
• Presentation via web page or email or
• Possibly distill the information.

10
Filtering Agent Architecture
News Server
Web Browser
Indexed Articles
User profiles
web
Indexing Engine
Other information sources
Filtering Agent
(from Caglayan)
11
Newshound (www.newshound.com)

• Delivers news each day by email or web.
• Searches San Jose Mercury News plus selected
  newswire sources.
• User defines profile(s)
• profile name
• required terms, possible terms, terms to exclude
• selectivity value (1..100)
• full email address/ account details
• Uses Verity Topics indexing engine.

12
Knowledge in Filtering Agents

• Newshound works with user-specified terms.
• Richer representations of interests ?
• Ideally a filtering agent could adapt
  automatically to users changing interests
• explicit user feedback (e.g. Firefly) or
• monitor user behaviour e.g.
• ...

13
Notification Agents

• Notify user of specified (internet) events
• update of a specified url
• Mind-It (www.netmind.com/html/url-minder.html)
• addition of url to a specified Yahoo subject
  category
• change in the results of a specified search query
• Server-based or desktop-based. Server -
• reduces network and processing
• can run same checks for multiple clients
• Neednt retrieve whole pages, just header.

14
Autonomy Agentware (www.agentware.com)

• See printout of web pages.
• Aimed at companys knowledge workers.
• Knowledge mgmt., classification,
• Serves as filtering agent , notification agent,
  ...
• Works with internet, intranet, Lotus Notes, SQL
  databases, ...

15
Autonomy (cont.)

• Filtering profiles based on
• natural language specification
• concepts mentioned in users own work.
• Advanced concept-matching tools.
• Can distinguish jobs in jobs on offer from
  Steven Jobs
• Adapts. How ?

16
Sharing Agents

• Agents which arrange for appropriate people in a
  group/organisation to get newly supplied
  information.
• Overlap with filtering and notification.
• See Jasper paper.
• System may assume some homogeneity in user
  interests.

17
Intelligent Browsing agents

• Suggest links to the browsing user.
• WebWatcher (server-based, internet agent)
• learns from thousands of other users visits and
  recommendations
• Letizia (client-based, desktop agent)
• collects info about users browsing habits and
  anticipates additional items of interest
• combines
• filtering - removes irrelevant material from
  search results and
• retrieval - forms queries and runs them during
  idle time.

18
Access to Structured Info

• Most internet info agents gather unstructured
  textual information.
• FAQfinder deals in semi-structured info
• There is also a need for agents that retrieve
  info from structured databases
• weather maps, stock quotes, flight timetables
• A full internet agent would know how to find
  out about all the internet databases and services
  and how to use them.

19
Other Internet Service Agents

• Job finding
• keep a watch for job ads that match criteria
• Bargain finding
• agent knows about vendors and their interfaces
• Financial
• gather uptodate info on stock prices, etc.
• Entertainment (Firefly - social info filtering)
• find clusters of people with the same tastes and
  share recommendations for music, films, amongst
  them.

20
The future ?Supply and Demand agents

• Demand agents request information on behalf of
  users (reactive or proactive)
• Supply agents provide information that matches
  requests from demand agents
• different source agents will cover different
  subject areas or kinds of info or
• they may compete on quality, reliability of info
  or on ability to match request precisely or
• see Brown et al, 1995

21
Issues

• Learning
• Modelling of info sources, etc.
• Modelling users
• ...

22
Info Agents in Intranets

• Intellignt browsing agents are particulalry
  promising in intranet environments where users
  interests are easier to anticipate and the
  information is more limited. (OLeary)
• ContactFinder connects queries posted to an
  intranet with employees perceived by
  ContactFinder to have relevant expertise (aka
  Expertise finders).

23
Intranet Agents (cont)

• Ideally an intranet agent would handle natural
  language queries like
• for the current month, which of the products have
  sales above budget ?

24
Exercise

• Design a (feasible) intelligent agent for your
  own use or for an information worker or ...
• What functions would it have ?
• How autonomous would it be ?
• What knowledge and intelligence would it need ?
  Where would it get it ? Could it learn ?
• Would it be mobile ?
• .

25
The future - CIG Searchbots

• CIG cooperative information gathering
• http//dis.cs.umass.edu/research/searchbots.html
• http//dis.cs.umass.edu/research/big/big.html
• Experimental/research system
• multi-agent approach to information retrieval

26
Characteristics of CIG search

• Goal-directed approach to searching
• as opposed to keyword driven
• e.g. find (in less than 10 minutes) information
  on word processors for a Mac, costing less than
  200 pounds, with these other characteristics ...
• multi-agent and parallel search
• one agent uses a traditional search engine,
  another goes to a software database, another
  looks for relevant Usenet news archives.

27
Characteristics (cont)

• inference based on gathered info drives further
  search
• e.g. names of products extracted in the first
  phase might be searched for in company product
  web pages, for desirable features, and then in
  review literature for objective assessments.
• knowledge-rich retrieval
• uses domain knowledge (e.g. about choosing
  software products)

28
Characteristics (cont)

• adaptive and dynamic search strategies
• search strategy develops as information becomes
  available.
• active search
• not off-line index-based search like search
  engines do
• assimilation of gathered information
• not just a list of urls, but (ideally) a report
  on the options and their realtive merits.

29
CIG draws on Research Areas

• problem-solving and planning
• deciding how to satisfy the goal
• adhering to resource (time, cost) constraints
• dynamically - as info becomes available
• scheduling
• how to order and interleave the actions
• multi-agent systems
• how to get agents to work together
• how to divide a problem among agents

30
Research ...

• information extraction
• e.g. finding products, producers, suppliers,
  features, cost, etc. in text
• interpretation
• creating information from raw data
• e.g. building a model (picture) of the range of
  WP packages available from a variety of sources.