Learning Qualitative Models

1
Learning Qualitative Models Ivan Bratko, Dorian
Suc Presented by Cem Dilmegani FEEL FREE
TO ASK QUESTIONS DURING PRESENTATION
2
Summary

• Understand QUIN algorithm
• Explore the Crane Example
• Analyze Learning Models expressed as QDEs
• GENMODEL by Coiera
• QSI by Say and Kuru
• QOPH by Coghill et Al.
• ILP Systems
• Conclusion
• Applications
• Further Progress

3
Modeling

• Modeling is complex
• Modeling requires creativity
• Solution Use machine learning algorithms for
  modeling

4
Modeling

• Modeling is complex
• Modeling requires creativity
• Solution Use machine learning algorithms for
  modeling

5
Learning

• examples hypothesis

learning
Hypothesis
examples
6
Decision Tree
7
Decision Tree Algorithm
8
QUIN (QUalitative INduction)

• Looks for qualitative patterns in quantitative
  data
• Uses so-called qualitative trees

9
Qualitative tree

• The splits define a partition of the attribute
  space into areas with common qualitative
  behaviour of the class variable
• Qualitatively constrained functions (QCFs) in
  leaves define qualitative constraints on the
  class variable

10
Qualitatively constrained functions (QCFs)

• The qualitative constraint given by the sign only
  states that when the i-th attribute increases,
  the QCF will also change in the direction
  specified in M, barring other changes.

11
Qualitative Tree Example
12
Explanation of Algorithm(Leaf Level)

• Minimal cost QCF is sought
• Cost M(inconsistencies or ambiguities between
  dataset and QCF)

13
Consistency

• A QCV (Qualitative Change Vector) is consistent
  with a QCF if either a) class qualitative change
  is zero b) all attributes QCF-predictions are
  zero or c) there exists an attribute whose QCF
  prediction is equal to the class' qualitative
  change
• ZM,-(X,Y)
• a) no change (inc,dec)
• a) no change (inc,inc)
• b) (no change, no change)
• c) inc (inc, dec)

14
Ambiguity

• A qualitative ambiguity appears a) when there
  exist both positive and negative QCF-predictions
  b) whenever all QCF-predictions are 0.
• ZM,-(X,Y)
• a) (inc,inc)
• b) (no change, no change)

15
Ambiguity-Inconsistency
16
Explanation of Algorithm

• Start with QCF that minimizes cost in one
  attribute and then use error-cost to refine the
  current QCF with another attribute
• Tree Level algorithm QUIN chooses best split by
  comparing the partitions of the examples it
  generates for every possible split, it splits
  the examples into 2 subsets (according to the
  split), finds the minimal cost QCF in both
  subsets and selects the split which minimizes the
  tree error cost. This goes on until, a specified
  error bound is reached.

17
Qualitative Reverse Engineering

• In the industry, there exists library of designs
  and corresponding simulation models which are not
  well documented
• We may have to reverse engineer complex
  simulations to understand how the simulation
  functions.
• Similar to QSI

18
Crane Simulation
19
QUIN Approach

• Looks counterintuitive?
• Yes, but it outperforms straightforward
  transformations of quantitative data to
  quantitative model, like regression

20
Identification of Operator's Skill

• Can't be learnt from operator verbally (Bratko
  and Urbancic 1999)
• Skill is manifested in operator's actions, QUIN
  is better at explaining those skills than
  quantitative models

21
Comparison of 2 operators

• S (slow) L
  (adventurous)

22
Explanation of S's Strategy

• At the beginning V increases as X increases (load
  behind crane)
• Later, V decreases as X increases (load gradually
  moves ahead of crane)
• V increases as the angle increases (crane catches
  up with the load

23
GENMODEL by Coiera

• QSI without hidden variables
• Algorithm
• Construct all possible constraints using all
  observed variables
• Evaluate all constraints
• Retain those constraints that are satisfied by
  all states, discard all other
• The retained constraints are your model

24
GENMODEL by Coiera

• Limitations
• Assumes that all variables are observed
• Biased towards the most specific models
  (overfitting)
• Does not support operating regions

25
QSI by Say and Kuru

• Explained last week
• Algorithm
• Starts like GENMODEL
• Constructs new variables if needed
• Limitations
• Biased towards the most specific model

26
Negative Examples

• Consider U-Tube Example
• Conservation of water until the second tube
  bursts or overflows
• There can not be negative amounts of water in a
  container
• Evaporation?

27
Inductive Logic Programming (ILP)

• ILP is a machine learning approach which uses
  techniques of logic programming.
• From a database of facts which are divided into
  positive and negative examples, an ILP system
  tries to derive a logic program that proves all
  the positive and none of the negative examples.

28
Inductive Logic Programming (ILP)

• Advantages
• No need to create a new program, uses established
  framework
• Hidden variables are introduced
• Can learn models with multiple operating regions
  as well

29
Applications

• German car manufacturer simplified their wheel
  suspension system with QUIN
• Induction of patient-specific models from
  patients' measured cardio vascular signals using
  GENMODEL
• An ILP based learning system (QuMAS) learnt the
  electrical system of the heart and is able to
  explain many types of cardiac arrhythmias

30
Suggestions for Further Progress

• Better methods for transforming numerical data
  into qualitative data
• Deeper study of principles or heuristics
  associated with the discovery of hidden variables
• More effective use of general ILP techniques.

31
Sources

• Dorian Suc, Ivan Bratko Qualitative Induction
• Ethem Alpaydin Introduction to Machine Learning
  MIT Press
• Wikipedia

32
Any Questions?

• ???????
• ?? ??
• ?? ??
• ??
• ??
• ??
• ??
• ??
• ??