Transcript from the lecture on IJCAI 99 in powerpoint format

Artificial Intelligence
AI in 1999: IJCAI 99
Ian Gent
[email protected]
Artificial Intelligence
AI in 1999
Part I :
Part II:
Part III:
Practical 1: Imitation Game
AI in 1999: IJCAI 99
Case based reasoning
Practical 1: The Turing Test
 Write a program to play the imitation game
 Some practical stuff:
 This is practical 1 of 2.
 Each will carry equal weight, I.e. 10% of total credit
 You may use any implementation language you wish
 Deadline(s) are negotiable
 to be decided this week
3
Practical 1: The Turing Test
 Write a program to play the imitation game
 Aim:
 to give practical experience in implementing an AI system
for the most famous AI problem
 Objectives:
 after completing the practical, you should have:
 implemented a dialogue system for conversation on a topic of you
choice
 gained an appreciation of some of the basic techniques necessary
 realised some of the possibilities and limitations of dialogue
systems
4
Some techniques you might use
 Pattern matching:
 my boyfriend made me … -> your boyfriend made you …
 I/me/my … -> you/you/your …
 Keyword identification & response
 my mother said …. -> tell me more about your family
 Deliberate errors
 34957 + 70764  105621
 mistypings
 Non sequiturs
 “Life is like a tin of sardines. You’re always looking for the
key”
5
Some pointers
 How to pass the Turing test by cheating
 Jason Hutchens, available on Course web pages
 Weizenbaum’s original paper on Eliza
 Comms ACM 1968
6
Your task




Choose a domain of discourse, e.g. Harry Potter
Implement a system to converse on this subject
Submit your program code, report, two dialogues
Program code
 in any language you wish
 I need an executable version to converse with
 e.g. via Web interface, PC/Mac executable, Unix
executable on a machine I can access
 consult me beforehand if in doubt
7
Your task
 Report
 A summary of the main techniques used and how they
work in your system
 a critical appreciation the main strengths and weaknesses
of your system
 (at least) Two Dialogues
 at least one dialogue with yourself
 to allow you to show off your system at its best
 at least one dialogue with another automated system
 e.g. Eliza on the web, a colleague’s system
8
What I am looking for
 A functioning program
 using appropriate technique(s) for playing the imitation
game
 need not have thousands of canned phrases
 need not be world standard
 should illustrate understanding of how to write programs to
play the imitation game
 A report summarising what you have done
 should be a minor part of the work for the practical
 no set word limit but probably just a few pages
 Some illustrative dialogues
 illustrating techniques and points in your report
9
IJCAI 99
 IJCAI 99 in Stockholm, Sweden, August 1999
 associated events such as workshops tutorial #
 IJCAI = International Joint Conference on AI
 leading AI conference
 every two years, odd years
 started in 1969
 other main conferences are AAAI, ECAI
 American Association for AI, five out of six years (really)
 European Conference on AI, even years
10
Topics at IJCAI 99, Volume 1








Automated Reasoning (32 papers)
Case Based Reasoning (6)
Papers responding to IJCAI-97 challenges (10)
Cognitive Modelling (8)
Constraint Satisfaction (12, should’ve been 13)
Distributed AI (12)
Computer Game Playing (4)
Knowledge Based Applications (9)
11
Topics at IJCAI 99, Volume 2









Machine Learning (29 papers)
Natural Language Processing (11)
Planning and Scheduling (13)
Qualitative Reasoning and Diagnosis (12)
Robotics and Perception (7)
Search (8)
Software Agents (3)
Temporal Reasoning (3)
Uncertainty and Probabilistic Reasoning (16)
12
IJCAI 99
 Every published paper passes peer review process
 usually three experts review paper
 programme committee selects best papers from these
 A co-operative effort …





37 members of the programme committee
400 reviewers
195 papers published
only 26% of total submissions
such a high standard that my submission was rejected!
 The state of the art of AI research in winter 98/99
13
Two Best Papers
 Two papers were selected by the P.C. as best
 IJCAI best paper awards always a bit of a lottery
 “A distributed case-based reasoning application for
engineering sales support”
 Ian Watson, Dan Gardingen
 “Learning in Natural Language”
 Dan Roth
 I will talk about Watson & Gardingen’s paper
 much more readable than Roth’s
 illustrates Case based reasoning, another area of AI
14
Distributed case based …
 Ian Watson,
 AI-CBR, University of Salford
 Dan Gardingen,
 Western Air Ltd, Fremantle, Australia
 “A distributed case-based reasoning application for
engineering sales support”
 Proceedings of IJCAI-99, pages 600-605
 A $32,000 project over 6 months to trial system
 Eventually fielded, $127,000 in Pentium notebooks
 Company estimates system made it $476,000 in 1st year
15
Distributed case based …
 Sales engineers distributed around Australia
 Quoting for Air conditioning/Heating systems
 Each quotation may be complicated
 sales engineers not qualified to quote
 fax details to central company
 wait for central engineers to supply quotation
 Company previously used database of past installations
 hard for sales staff to find similar quotes
 How could Case based reasoning system help?
16
Case based reasoning
 a problem solving strategy using existing cases
 to automate ‘knowledge reuse’
 assume previous cases have been correctly dealt with
 cases might have been addressed by humans
 associate with a case a set of feature-value pairs
 together form a unique index for the case
 possibly weight features with importance score
 use existing case database to help with new cases
 calculate index of new case
 find some number of the ‘closest’ cases
 use these to help treat new case
17
Cases for HVAC
 HVAC = heating, ventilation, air conditioning
 Each case contains 60 fields for retrieval
 plus further fields describing installation
 plus links to ftp area for download
 Aim is to find some ‘nearest neighbour’ cases
 From these, sales staff can look at a small number
of similar cases, and adapt quotes
 Quotes confirmed at central site
 In trial, expertise of central engineers never used
 just for checking quotes that the sales staff proposed
 One benefit is saving in central experts time
18
Finding similar cases
 Finding the similar cases is not rocket science
 Remember, aim is to find a few similar cases
 can be used by field staff as basis for new quote
 want a manageable number (e.g. 20)
 Main technique is to relax values of features
 e.g. “item Athol_B23” becomes “T31_fan_coil”
 where Athol_B23 is one specific type of T31_fan_coil
 allows retrieval of installations using other types
 e.g. “temperature = 65 F” becomes “60F < T < 65F”
 Knowledge engineering used to find relaxations
 e.g. use of domain experts to advise on suitable relaxations
19
Distributed reasoning...
 System was distributed using Java & XML
 Server uses relaxation to produce reasonable
number of items, e.g. a few hundred
 Pushed to client side applet via XML
 runs simple nearest neighbour algorithm to find closest set
 Simply minimise similarity measure
 i f(Ti,Si) wi
 where summation over features i
• f(Ti,Si) difference measure on feature i between cases S, T
• wi is weight of feature i
 obtain full details of closest set by ftp
20
How did this win the lottery?
 Not exactly rocket science
 I’ve almost presented all the technical details already
 Web, Java, and HTML in paper can’t have hurt it!
 Shows a real world application
 saved a company some real money
 Shows maturity of an AI technique
 here, case based reasoning
 fielded good application in 6 months for only $32,000
21