Artificial Intelligence its not just for breakfast anymore

1
Artificial Intelligence(its not just for
breakfast anymore)

• 1 May 2003
• Anthony (Tony) J. Grichnik
• Strategy Technology Manager
• Caterpillar Inc.
• grichaj_at_cat.com

2
Overview

• Technologies
• Expert Systems and the Bagel button
• Fuzzy Logic and your thermostat
• Inference, clustering and junk mail
• Neural Networks for the organic gardener
• Genetic Algorithms and your 401(k)
• Perspective
• Biggerbetterfastermore vs. Occams Razor
• The Knowledge-Based Enterprise model
• Tomorrow
• Grids
• Hybrids

3
Expert Systems and the Bagel button

• What is it?
• A classical rule structure (if-then-else) used to
  describe and select an outcome from a series of
  choices
• Advantages
• Easily understood and maintained
• Mimics the results of expert learning so the
  experts knowledge can be in many places at once
• Drawbacks
• They dont learn like an expert.
• Knowledge becomes obsolete.
• Rules dont adapt to change.
• Theyre impractical for complex problems.
• Geometric growth (2n) is an issue.
• Theyre vulnerable to perception.
• How big is your bagel?
• Where to find them
• Microwaves, car parts store, makeup counter,
  phone support lines

4
Fuzzy Logic and your thermostat
Goal 68 oF at 500 PM

• What is it?
• Typically an Expert System or control structure
  modified with trueness and falseness
  represented mathematically
• Advantages
• Fundamentally similar to expert systems
• More adaptable than a hard expert system
• Drawbacks
• Takes general expert system issues
• They dont learn like an expert.
• Theyre impractical for complex problems.
• Theyre vulnerable to perception (to a lesser
  degree).
• and adds a few more!
• Geometric growth actually gets worse (Dn) instead
  of better when fuzziness is added.
• Converts logical problems into computational ones
  (which is bad for the typical CPU).
• Where to find them
• Smart thermostats, adaptive automatic
  transmissions in cars, some dishwashers

Traditional thermostat
Fuzzy thermostat
5
Inference, clustering and junk mail

• What are these?
• General AI terms for processes which take large
  numbers of attributes and reduce them to smaller
  numbers of choices or sets
• Advantages
• Determines effective structures on the fly,
  making them adaptable yet fully logical.
• Often expresses complex problems more efficiently
  than human programmers.
• Drawbacks
• Dependent on stable input relationships with
  contextual consistency
• Father gt Parent? or priest?
• Short gt Height? or electricity? or
  length? or finances?
• Where to find them
• Diagnostic support, advertisement profiling,
  survey interpretation, law enforcement
  (controversial), intelligence gathering

6
Neural Networks for the organic gardener

• What are they?
• In general, neural networks are mathematical
  models of neurological functions, usually
  mammalian or human.
• NNs usually fall into two categories
  synthesizers and classifiers.
• Advantages
• The natural NN structures can represent very
  complex relationships in comparatively simple
  mathematical ways.
• Highly adaptive solution for many problems.
• Drawbacks
• NNs can rarely explain why they do what they do
  in logical terms.
• This can make them difficult to debug when they
  do not behave as desired.
• NNs start off dumb and must learn by
  experience, which can be time consuming.
• Where to find them
• OCR software, speech recognition, financial
  applications, fraud detection

The Slugbot (University of The West of England
Bristol)
7
Genetic Algorithms and your 401(k)

• What are they?
• GAs are search algorithms that use natural
  principles (reproduction, selection, mutation) to
  build solutions.
• Advantages
• GA performance is almost unequaled for complex
  searches.
• Highly adaptive solution for many problems.
• Drawbacks
• GAs always need a goal function to determine
  which solutions survive and which ones die.
• The goal function elements must (ideally) remain
  consistent during operation.
• Not all problems have an obvious goal function.
• GA / system interactions can be complex.
• Especially true if actions are being taken based
  on the GAs output.
• Where to find them
• Portfolio balancing, the search for new materials
  or medicines, job scheduling

8
Overview

• Technologies
• Expert Systems and the Bagel button
• Fuzzy Logic and your thermostat
• Inference, clustering and junk mail
• Neural Networks for the organic gardener
• Genetic Algorithms and your 401(k)
• Perspective
• Biggerbetterfastermore vs. Occams Razor
• The Knowledge-Based Enterprise model
• Tomorrow
• Grids
• Hybrids

9
Biggerbetterfastermore vs. Occams Razor
Biggerbetterfastermore
Occams Razor

• Corollary to Moores Law
• CPU performance roughly doubles every 18 months
• Expendable computing
• All hardware is perishable
• All software is transient
• Reuse of either is uneconomical(so either reduce
  or recycle)
• Perpetual Restriction Cycle Theory
• Memory (active storage)
• Disk (long-term storage)
• Processor (active manipulation)
• Network (passive manipulation)

• Occams Razor (Parsimony theory)
• One should not increase, beyond what is
  necessary, the number of entities required to
  explain anything.
• Translation That which describes simplest,
  describes best.
• KISS Method
• Keep It Simple, Stupid!
• Unrestricted Computing Theory
• If you think youre restricted, either
• Youre solving the wrong problem.
• Youre solving the right problem the wrong way.

10
The Knowledge-based Enterprise Model

• To what degree can we act autonomously on what we
  know?
• How can we communicate what we know to people and
  systems that need to know it?
• How can we capture the learning process that
  converts information into knowledge?
• Do the correct associations exist between pieces
  of information?
• Is the data being correctly organized into useful
  information?
• Are we consistently collecting the right data the
  right way?

Take Action!
Knowledge
Information
Collect Data
11
Overview

• Technologies
• Expert Systems and the Bagel button
• Fuzzy Logic and your thermostat
• Inference, clustering and junk mail
• Neural Networks for the organic gardener
• Genetic Algorithms and your 401(k)
• Perspective
• Biggerbetterfastermore vs. Occams Razor
• The Knowledge-Based Enterprise model
• Tomorrow
• Grids
• Hybrids

12
Grid Computing
Grid is a type of parallel and distributed
system that enables the sharing, selection, and
aggregation of resources distributed across
multiple administrative domains based on their
(resources) availability, capability,
performance, cost, and users' quality-of-service
requirements. Rajkumar Buyya, University of
Melborne, GRIDS Lab (gridbus project)
13
Hybrids
What happens when we can implant artificial
neural networks in real brains?
What happens if we fuzzify a GAs goal function?
Can computers really be creative? If so, what
would they do? Why do we do what we do?
Why cant a GA evolve an expert?
Will AI and Human Intelligence mesh or clash in
the future?
What about a NN trained from inference or
clustering?
Any sufficiently advanced technology is
indistinguishable from magic. Arthur C. Clarke
Should any AI be solely responsible for any
decision?
14
Artificial Intelligence(its not just for
breakfast anymore)

• 1 May 2003
• Anthony (Tony) J. Grichnik
• Strategy Technology Manager
• Caterpillar Inc.
• grichaj_at_cat.com