COSC 4355 and 5355 Expert Systems

1
COSC 4355 and 5355Expert Systems

• Introduction to Expert Systems (Part 2)
• Dr. Lappoon R. Tang

2
Overview

• Basic concepts in implementing an expert system
  languages, shells, and tools
• Diving another level down what elements do we
  have in an expert system?
• Treating an expert system as a production system

3
Readings

• ESPP
• Section 1.8 to Section 1.10

4
Some basic concerns in constructing an expert
system

• Since an expert system is a piece of software
  (remember an expert was somehow turned into a
  system?), what would be a good programming
  language for implementing the system?
• Are there tools that can facilitate development
  of an expert system?
• Are there tools that can be used to develop an
  expert system?

5
Expert system languages?
6
Tools for making life easier in creating expert
systems?

• Tools here refer to utility programs that
  facilitate development of an expert system (as
  opposed to those that are actually used to
  develop one)
• Debugger
• Text and graphics editors
• File management systems
• In some cases, even code generator

7
What about tools that are used to make expert
systems?

• Tools that are used directly to make an expert
  system is called an expert system shell
• A special tool that allows a knowledge engineer
  to encode knowledge in a programming-language
  like formalism
• Notice
• Expert system shells are not general purpose
  programming languages they are more like
  language like tools for specifying knowledge
• However, some programming languages do facilitate
  creation of expert system shells (e.g. PROLOG)
• Examples
• CLIPS
• Any expert system shell created by PROLOG
• EMYCIN

8
Review Outline of an expert system (dont tell
me u forgot -p)
Knowledge Base
Someone who needs to solve a problem
Facts
Rules
Expertise (e.g. decisions, answers)
Inference Engine
BUT He doesnt have the knowledge!! ?
Expert System
9
A much more in-depth outline of an expert system ?
10
The Knowledge Base

• Contains production rules IF THEN rules

(A sample rule from MYCIN translated into
English) IF 1. the site of the culture is
blood 2. the identity of the organism is not
known with certainty 3. the stain of the
organism is gramneg 4. the morphology of the
organism is rod 5. the patient has been
seriously burned THEN Conclusion the identity
of the organism is pseudomonas (CF 0.4)
LHS Antecedents
RHS Consequence

• MYCIN has around 600 such rules ?

11
Knowledge Acquisition

• The knowledge acquisition (KA) facility allows
  the expert system to induce new rules from
  observed data
• An optional feature in an expert system (in fact
  most expert systems dont have this capability)
• Most KA facilities, if not all, are built on
  (rule based) machine learning algorithms such as
  ID3, C4.5 (latest version C5.1)

12
The working memory

• This is where facts (related to the problem we
  need to solve) are stored in the expert systems
• Example Medical diagnosis
• Fact 1 the patient has a running nose
• Fact 2 the patient does not have a sore throat
• Fact 3 the patient keeps coughing

13
The inference engine

• Using facts stored in the working memory, the
  inference engine checks which rules have all
  conditions (i.e. antecedents) satisfied
• Rules with their antecedents satisfied (aka
  activations) are stored in the agenda the rule
  with the highest priority is used first
• If an activation is applied, the RHS (i.e. the
  consequence) becomes true or an action is taken
  if the RHS is an action. Notice that new facts
  are added back to the working memory.
• Basically, in this case, the inference engine
  engages in a recognize-act / situation-response
  cycle until the problem is solved (e.g. a
  diagnosis is found, a computer is configured)
• Note an activation that has been applied is
  removed from the agenda to avoid unnecessary
  repetition
• Generally, there are two ways to perform rule
  based inference
• Forward chaining use all the facts to find out
  which rules have become activated, apply the rule
  with highest priority, add new facts derived back
  to working memory
• Backward chaining start with the RHS of a rule,
  see if the conditions on the LHS are all true (in
  the working memory), if so, apply the rule and
  add new facts derived back to the working memory

14
Explanation facility

• This component of an expert systems allows a user
  to
• 1) find out how a conclusion is arrived (e.g. a
  diagnosis) or a solution is produced by asking
  why questions to the systems
• 2) explore alternative ending (i.e. alternative
  solution or diagnosis) by asking what if
  questions (like what if we had this alternative
  piece of evidence )

15
The user interface

• It performs mainly two functions
• 1) Allows a user to control (at the top level)
  how the expert system shell works by giving
  commands (e.g. how inference should be performed,
  backward chaining or forward chaining?)
• 2) Displays information relevant for solving the
  problem at hand
• Example Warning messages
• Example Relevant information about the problem
  itself (e.g. current status of CPU flags)

16
Issues with ordering does ordering of rules
matter?

• It turns out the order of application of all
  activations (i.e. rules that are applicable) does
  matter

(Suppose say both rules are applicable) IF there
is a fire THEN leave IF my clothes are on fire
THEN put out the fire

• Suppose you and your friend are escaping a fire
  but unfortunately your clothes catch on a fire in
  the process. Both rules will apply but obviously
  the 2nd rule has a much higher priority ?

17
Solution 1 Markov Algorithm

• Markov (Andrey Andreyevich Markov Jr., 1903
  1979) came up with the idea of giving priority to
  rules so that all applicable production rules are
  NOT un-ordered but rather order
• Example
• 1) car wont start ? check gas
• 2) car wont start ? check battery

18
Solution 1 Markov Algorithm

• Markov algorithm
• 1) If the rule with the highest priority applies,
  apply it
• 2) If not, apply the next rule with the highest
  priority (i.e. the one with 2nd highest priority)
• 3) Keep performing 1) and 2) until either
• a) the computer blows up (just a joke ?)
• b) no rule is applicable
• c) the problem is solved

19
Solution 2 Rete Algorithm

• Unfortunately, the Markov algorithm requires one
  to sequentially go through a list of rules to see
  which one applies
• What if the list of rules become huge?
• Answer we would rather check out a DVD instead
  of waiting for the PC to blow up -p
• Lesson Issue of efficiency matters!

20
Solution 2 Rete Algorithm

• Basically, Charles L. Forgy at CMU recognized
  this problem in the late 70s and came up with
  the Rete algorithm to solve the problem of
  efficiency associated with Markov algorithm
• Rete algorithm is designed with two observations
  in mind
• Temporal redundancy application of a rule
  usually changes only a few facts AND only a few
  rules are affected by those new facts (or
  deletion of old facts)
• Structural similarity The same antecedent often
  appears in the LHS of more than one rule
• Idea use a B tree to organize the rules into a
  data structure that support fast searching of
  rules with matching antecedents AND look only for
  changes in matches this results in tremendous
  speed up

21
Open problems

• Problem 1 Can one use machine learning to learn
  priorities for a set of production rules if
  they were initially unordered?
• Problem 2 If we combine the same antecedent
  (e.g. car wont start) that appears in different
  production rules into a single node, we
  essentially have a network structure. Can one
  learn such a structure (i.e. essentially learning
  the rules) via machine learning for knowledge
  acquisition?
• Hint learning network structure for neural
  networks ?