Intelligent Systems CSCI 6501 Dr. D. Riordan

1
Intelligent SystemsCSCI 6501Dr. D. Riordan

• Pradeep Monga (B00342080)
• Satwant Sandhu (B00201045)

2

FINAL REPORT

• Structure for Credit-Apportionment Problem in
  Rule Based Systems

3

• Overview

• In this Project we have worked with Credit -
  Apportionment Problem in Rule Based Systems.
• We have implemented a hybrid expert system called
  GAMBLE (Genetic Algorithm Based Machine Learning
  Expert).

4
Contd

• The Credit-Apportionment process provides a
  formal basis for the problem analysis and
  algorithm design.
• It includes
• System Environment sub - model which provides
  integrated
• view about the payoff to, as well as
  the external and internal
• aspects of rule based system.
• Principles of Usefulness, which define the
  usefulness of rule
• actions.
• Definitions of the Credit-Apportionment problem
  which guides
• the algorithm synthesis.

5
Problem Formulation

• Credit-Apportionment problem can be formulated
  as-
• Local level problem
• Estimation of the inherent usefulness values
  in a particular context.
• Global level problem
• As approximation to the inherent usefulness
  functions ( ) from the payoffs.

6

• GAMBLE

• This System is used for students who are
  seeking admission into an engineering institute
  after clearing the entrance examination. The
  student is advised by this system as to which
  branch would be most suited for him, with the
  help of an algorithm used for branch selection.

7
Implementation

• We have following information
• No. of Seats available in each Branch,
• Thresholds of parameters (like logic, memory,
  aesthetic-sense, adaptability) of each Branch,
• Records of students already admitted (includes
  scores, averages, standard deductions, classifier
  strings and their weights.
• On starting execution, system prompts for
• Student Name
• Logic score
• Memory score
• Aesthetic-sense score
• Adaptability score

8
Branches and Parameters
Table showing sample threshold values of branches
in different parameters
9
Calculation of Branch Aptitude Total

• BAT of each branch for the candidate is
  calculated as follows,
• BAT ? p LG,IM (XBP SBT) / XBT
  
• Where,
• B Subscript for a particular
  branch,
• P Subscript for a particular
  parameter,
• XBP Weight of parameter P for branch B,
• SBT Score obtained by student in a particular
  parameter,
• XBT Total of the weighted parameters.

10
What GAMBLE does?

• GAMBLE shows BAT of each Branch and suggests the
  most suitable Branch for the candidate.
• It shows the branches in which student is
  ineligible to seek admission and prompts to
  choose among the branches in which the candidate
  is eligible (i.e. in which he/she has more than
  minimum marks and seats are available).
• Student is granted admission, if seats available
  in the chosen branch.
• Following information is updated in the database
  
• No. of seats available,
• Record of new student is added,
• and Learning process is started.

11
How are classifier strings generated?
In this case the generated string is
10where 1 is for logic, is for
memory, 0 is for Aesthetic-sense is for
Adaptability
Rules 1 if input gt Avg Sd/2 0 if input lt Avg
Sd/2 if Avg Sd/2 lt input lt Avg if Avg gt
input gt Avg Sd/2
12
More insight..
Performance
13
Comparison between two classifiers

• Two classifiers do not match if one classifier
  has 0 at a position and other has 1 at the same
  position or vice-versa, else in all other cases
  classifiers match.
• or are fuzzy variables and match with
  any value in other classifier.
• For e.g. 001 and 10 Match
• 10 and 1100 Match
• 10 and 00 - Unmatched

14
Learning Process

• Threshold classifier strings are compared with
  the classifier strings of all the students that
  have been admitted in the past.
• The strengths of classifiers that matched with
  threshold classifier are increased while the
  strengths of classifiers that didnt matched are
  reduced.
• One winning classifier is chosen randomly amongst
  the classifiers that matched and had
  comparatively higher strengths.
• Strengths of those classifiers is again increased
  by some percentage that were equally competent
  but couldnt win. They are rewarded so that they
  have better chance in future.
• The parameter values of winning classifier are
  ascertained and they are made the new thresholds.

15
Artificial Classifier generation

• In case, there is no match for threshold
  classifier in the records, the system is robust
  enough to handle the situation by unleashing the
  power of Genetic Algorithms. System uses a
  mechanism that implements the tripartite process
  of reproduction, crossover and mutation to
  produce the temporary classifiers.
• Fitness function
• If the incoming threshold message element is 1
  then the corresponding classifier element should
  not be 0 .
• if the incoming threshold message element is 0
  then the corresponding classifier element should
  not be 1.
• if the incoming threshold message element is 1
  then the corresponding parameter value is AVG
  d/8.
• if the incoming threshold message element is 0
  then the corresponding parameter value is AVG -
  d/8.
• In all other cases the parameter value is AVG.

16
Conclusions

• The learning mechanism is one of a clear
  candidate for a cognitive invariant in humans
  which involves the ability to acquire facts,
  skills and more abstract concepts.
• human learning aspects can be reproduced in a
  computer system by understanding the criteria by
  means of which humans learn.
• In the coming days and also in present
  situations, learning would tend to be more
  efficient than programming.
• An important aspect of student education has been
  covered in this report and the field is still
  open to make the system handle the effect of
  various other changes in the environment and its
  response towards them.

17
References

• 1 http//www.icce2001.org/cd/pdf/p14/IN002.pdf
• GAMBLE expert system , Credit apportionment
  process and Bucket Brigade Algorithm Indian
  Institute of Technology, Roorkee
• India 2001.
• 2 IEEE Transactions On Systems And
  Cybernatics, Framework for the
  Credit-Apportionment Process in Rule-Based
  Systems Vol 19, No 3, May/June 1989.
• 3 Bucket Brigade Performance 1 long sequences
  of classifiers in Genetic algorithms and their
  application proc 2nd int. conf on GA.
• J.Grefenstette, Ed. July 1987.
• 4 A study on apportionment of credits of fuzzy
  classifier system for knowledge acquisition
  of large scale systems Nakaoka, K. Furuhashi,
  T. Uchikawa, Y.Fuzzy Systems, 1994. IEEE World
  Congress on Computational Intelligence,
  Proceedings of the Third IEEE Conference on,
  26-29 June 1994

18
Contd..

• 5 Goldberg, David E., Genetic Algorithms in
  Search Optimization, and Machine
  Learning, Addison Wesley Longman, International
  Student
• Edition 1999.
• 6 J. Holland. Escaping brittleness the
  possibilities of general purpose
• learning algorithms applied to parallel ruled
  based systems. In
• R. Michalski, J Corbonell and T Mitchell,
  editors, Machine learning
• An Artificial intelligence approach, Morgan
  Kauffmann Publishers, Inc.
• Los Altos, CA., 1986

19

• Any Questions are welcome?