Automatically Integrating Multiple Rule Sets in a Distributedknowledge Environment

1
Automatically Integrating Multiple Rule Sets in a
Distributed-knowledge Environment
C.H. Wang, T.P. Hong, S.S. Tseng, C.M. Liao, IEEE
Trans on SMC, Part C, Vol. 28, No. 3, August 1998

• Teacher ???
• Student ???
• no M8702048
• date 6/11/99

2
Why use Distributed-knowledge Environment

• Developing a expert system requires knowledge
  base and its knowledge is often distributed among
  groups of a single experts.
• Acquiring and integrating multiple knowledge
  inputs form many experts or by various
  knowledge-acquisition technique thus plays an
  important role in building effective
  knowledge-based system.

3
Knowledge-integration methods problem

• Domain experts must intervene during integration
  to resolve conflicts and contradictions
• integration is time-domain(require weeks or
  months)
• The more knowledge sources consulted, the more
  difficult and complex the integration

4
Genetic knowledge-integration method

• Automatically combines multiple rule sets into
  one integration
• each rule set is encoded as a bit string and
  evaluated by an evaluation function
• Domain experts need not intervene in the
  integration process
• experimental results show that the propose
  approach can greatly improve the knowledge base

5
Knowledge integration

• All knowledge derived by knowledge-acquisition(K.A
  .) tools or induced by machine-learning(M.L.)
  methods
• two process
• encoding
• integration

6
Knowledge encoding

• Michigan approach encodes individual rules into
  fixed-length bit strings
• Pittsburgh approach encoding rule sets into
  variable-length bit strings, with each individual
  in the population representing a rule set
• the rule sets from different sources be
  translated into a uniform syntactical
  representation before being encoded

7
Encoding step

• Collect the features and possible values form the
  condition port of rule sets
• collect classes form the conclusion parts of the
  rule sets
• translate each rule into an intermediary
  representation that retains its essential syntax
  and semantics. Dummy tests are inserted into the
  condition part of the rule
• Each feature test is then encoded into a
  fixed-length binary string, so is the class
  pattern
• for each rule set, concatenate all its rule
  substrings. Different rule sets might be different

8
Example 1

• Class Adenoma, Meningioma
• Feature Location, Calcification, Edema
• Locationbrain surface, sellar, brain stem
• Calcificationnone, marginal vascular-like,
  lumpy
• Edemanone, lt2 cm, lt0.5hemisphere
• R1 if (Location sellar) and
  (Calcificationno) then class is Adenoma
• R2 if (Locationbrain surface) and (Edema lt
  2cm) then class is Meningioma

9
Cont.

• R1 if (Location sellar) and
  (Calcificationno) and (Edema no or Edema lt 2cm
  or Edema lt 0.5hemisphere) then class is Adenoma
• R2 if (Locationbrain surface) and (Edema lt
  2cm) and (Calcification no or Calcification
  marginal or Calcification vascular-like or
  Calcification lumpy) then class is Meningioma
• Rule Location Calcification Edema Class
• R1 010 1000 111 10
• R2 100 1111 010 01
• gt 010 1000 111 10 100 1111 010 01

10
Knowledge integration
Training instances,
11
Crossover and Mutation

• Crossover
• select a crossover point in one of the parents at
  random
• if point in rule boundary, the another must in
  boundary if in p bit to boundary, another must in
  p bits boundary
• cross the genes of the parents according to the
  point
• generate new offspring
• Mutation
• random change some elements in a selected rule
  set.

12
Fusion

• Redundancy
• IF d1 or d2 THEN d3
• IF d1 or d2 THEN d3
• subsumption
• IF d1 or d2 THEN d3
• IF d1 THEN d3

13
Fission

• misclassification
• IF d1 or d2 THEN d3
• IF d1 THEN d3 IF d2 THEN d3
• contradiction
• IF d1 THEN d2 and d3
• IF d1 THEN d2 IF d1 THEN d3

14
Experimental Result

• Goal six possible classes of brain tumors
• 504 cases (70)train set (30)test set
• Each rule consisting of twelve feature tests and
  a class pattern was encoded into a bit string 105
  bits long

15
cont.

• Crossover0.9 mutation0.04 fusion,
  fission0.01
• change different domain-specific operator rates.
• Fusion operations reduced the structural
  complexity of the resulting rule sets, but
  fission operations increased it
• fusion dont affect the accuracy of the resulting
  rule sets, but fission increase it.

16
Future investigations

• Knowledge source or actual instance may contain
  fuzzy information in the real world
• Many issues in the field of knowledge
  verification remain unresolved