Advantages of Approximate Reasoning

  KBS : Approximate Reasoning Motivation

‹

reasoning for real-world problems involves missing

  ‹ Motivation ‹ Fuzzy Logic

  knowledge, inexact knowledge, inconsistent facts or ‹

  Fuzzy Sets and Natural

  ‹ Objectives

  rules, and other sources of uncertainty Language

  ‹

  Approximate Reasoning ‹ Membership Functions ‹ while traditional logic in principle is capable of

  ‹ Variation of Reasoning with ‹ Linguistic Variables capturing and expressing these aspects, it is not

  Uncertainty very intuitive or practical

  ‹

  ‹ Important Concepts and Commonsense Reasoning

  Terms ‹ explicit introduction of predicates or functions

  ‹ Chapter Summary

  ‹ many expert systems have mechanisms to deal with uncertainty

  ‹

  sometimes introduced as ad-hoc measures, lacking a sound foundation Approximate Reasoning 1 Approximate Reasoning 2

  

Objectives Approximate Reasoning

  ‹ be familiar with various approaches to approximate reasoning

  ‹ inference of a possibly imprecise conclusion from

  ‹ understand the main concepts of fuzzy logic

  possibly imprecise premises ‹ fuzzy sets

  

‹

useful in many real-world situations

  ‹ linguistic variables

  ‹ one of the strategies used for “common sense” reasoning

  ‹ fuzzification, defuzzification

  ‹ fuzzy inference

  ‹

  frequently utilizes heuristics

  ‹ evaluate the suitability of fuzzy logic for specific tasks ‹ especially successful in some control applications

  ‹ application of methods to scenarios or tasks ‹ often used synonymously with fuzzy reasoning

  ‹ apply some principles to simple problems

  ‹ although formal foundations have been developed, some problems remain

  Approximate Reasoning 3 Approximate Reasoning 4 Approaches to Approximate Advantages of Approximate Reasoning Reasoning

  ‹ fuzzy logic

  ‹ reasoning based on possibly imprecise sentences

  ‹ common sense reasoning ‹ default reasoning ‹ allows the emulation of some reasoning strategies used by humans

  ‹

  in the absence of doubt, general rules (“defaults) are applied ‹ concise

  ‹ default logic, nonmonotonic logic, circumscription ‹

  can cover many aspects of a problem without explicit representation of the details ‹ analogical reasoning

  ‹

  conclusions are derived according to analogies to similar ‹ quick conclusions situations

  ‹

  can sometimes avoid lengthy inference chains Approximate Reasoning 5 Approximate Reasoning 6

  Problems of Approximate Reasoning Fuzzy Logic ‹ nonmonotonicity ‹ approach to a formal treatment of uncertainty

  ‹ inconsistencies in the knowledge base may arise as new

  

‹

relies on quantifying and reasoning through natural sentences are added language

  ‹

  sometimes remedied by truth maintenance systems

  ‹ linguistic variables

  ‹ ™ used to describe concepts with vague values semantic status of rules

  ‹ default rules often are false technically ‹ fuzzy qualifiers

  ‹ efficiency ™ a little, somewhat, fairly, very, really, extremely

  ‹

  although some decisions are quick, such systems can be

  ‹ fuzzy quantifiers

  very slow ™ almost never, rarely, often, frequently, usually, almost always ™ especially when truth maintenance is used ™ hardly any, few, many, most, almost all

  Approximate Reasoning 7 Approximate Reasoning 8

  Approximate Reasoning 9 Fuzzy Sets and Example

  Possibility Measure ‹ degree to which an individual element x is a potential member in the fuzzy set S Poss{x

  ‹

  probabilities p(X = 7) = 2*3/36 = 1/6 7 = 1+6 = 2+5 = 3+4

  ‹

  X is an integer in U = {2,3,4,5,6,7,8,9,19,11,12}

  ‹

  ‹ Example: rolling dice

  ‹ possibility refers to allowed values

‹

probability expresses expected occurrences of events

  Approximate Reasoning 12 Possibility vs.. Probability

  Poss(A ∧ B) = min(Poss(A),Poss(B)) ™ Poss(A ∨ B) = max(Poss(A),Poss(B))

  a popular one is based on minimum and maximum ™

  ‹ various rules are used ‹

  ∈S} ‹ combination of multiple premises with possibilities

  short medium tall Approximate Reasoning 11

  membership height (cm) 50 100 150 200 250

  1

  0.5

  membership height (cm) 50 100 150 200 250

  Approximate Reasoning 10 Fuzzy vs. Crisp Set

  values in between indicate how strongly an element is affiliated with the set

  ™

  0 means no, 1 means full membership

  ™ associates each element xi with a degree of membership in S:

  described through a membership function µ(s) : x → [0,1]

  ‹

  short medium tall ‹ categorization of elements xi into a set S

  1

  0.5

  possibilities Poss{X = 7} = 1 the same for all numbers in U Fuzzification Fuzzification Example

  2 x

  1

  2

  3

  4

  ‹ function f(x) = (x-1)

  ‹ the extension principle defines how a value, function

  1

  1

  4

  9 f(x) or set can be represented by a corresponding fuzzy membership function

  ‹ known samples for

  1

  2

  3

  4

  ‹ extends the known membership function of a subset to a

  membership function specific value, or a function, or the full set

  0.5

  1

  0.5 “about 2”

  “about 2” function f: X → Y membership function µ for a subset A ⊆ X ‹

  A membership function of

  1

  2

  3

  4 x f(“about 2”) extension µ ( f(x) ) = µ (x)

  f(A) A

  1

  4

  9 f(x)

  0.5

  1

  0.5 f (“about 2”)

  [Kasabov 1996] Approximate Reasoning 13 [Kasabov 1996] Approximate Reasoning 14

Defuzzification Fuzzy Logic Translation Rules

  ‹ describe how complex sentences are generated from

  ‹ converts a fuzzy output variable into a single-value elementary ones variable

  ‹ modification rules

  ‹ widely used methods are ‹ introduce a linguistic variable into a simple sentence

  ™

  e.g. “John is very tall” ‹ center of gravity (COG)

  ‹ composition rules

  ™ finds the geometrical center of the output variable ‹ combination of simple sentences through logical operators

  ‹ mean of maxima

  ™

  e.g. condition (if ... then), conjunction (and), disjunction (or)

  ™ calculates the mean of the maxima of the membership function

  ‹

  quantification rules ‹ use of linguistic variables with quantifiers

  ™

  e.g. most, many, almost all ‹

  qualification rules ‹ linguistic variables applied to truth, probability, possibility

  ™

  e.g. very true, very likely, almost impossible

  [Kasabov 1996] Approximate Reasoning 15 Approximate Reasoning 16

  Approximate Reasoning 17 Fuzzy Inference Methods

  0.3

  1

  0.45

  0.3

  0.44

  0.43

  0.42

  0.41

  0.3

  0.4

  0.7

  1

  0.5

  1 1 bad_cc good_cc

  1

  1

  8 C Credit

  9

  10

  7

  0.7

  6

  0.3

  0.7

  0.1 C Ratio

  4

  5

  1

  1

  1

  8 Decision

  9

  10

  7

  0.7

  6

  0.3

  4

  1

  0.3

  3

  0.7

  2

  1

  1

  1

  0.7 bad_cc good_cc

  1

  1

  5

  0.3

  ‹ how to combine evidence across fuzzy rules

  ‹ principles that allow the generation of new sentences from

  ‹ represented as a set of two rules with tables for fuzzy set

  Example Fuzzy Reasoning 1 ‹ bank loan decision case problem

  (X,Y) is R Y is max(F,R) Approximate Reasoning 19

  X is F F ⊂ G X is G X is F

  F, G, R are relations

  X,Y are elements

  ‹ compositional rule

  ‹ entailment principle

  ‹ examples

  existing ones ‹ the general logical inference rules (modus ponens, resolution, etc) are not directly applicable

  Approximate Reasoning 18 Fuzzy Inference Rules

  C Score, CRatio, CCredit, Decision ™ fuzzy values high score, low score, good_cc, bad_cc, good_cr, bad_cr, approve, disapprove Rule 1: If (CScore is high) and (CRatio is good_cr) and (CCredit is good_cc) then (Decision is approve)

  derived from Lukasiewicz logic

  ‹ integrated with fuzzy logic based on the qualification translation rules ™

  ‹ e.g. fuzzy qualifiers like very likely, not very likely, unlikely

  ‹ describes probabilities that are known only imprecisely

  ™ extension of binary logic to infinite-valued logic Fuzzy Probability

  ‹ formal foundation through Lukasiewicz logic

  also Max-Product inference and other rules

  ‹

  ™ implication according to Max-Min inference

  ‹ Poss(B|A) = min(1, (1 - Poss(A)+ Poss(B)))

  definitions ™ fuzzy variables

  Rule 2: If (CScore is low) and (CRatio is bad_cr) or (CCredit is bad_cc) then (Decision is disapprove )

  3

  0.7

  0.7

  2

  1

  1

  1

  1 low high

  1

  1

  1

  150 185 200 195 190 C Score

  180

  [Kasabov 1996] Approximate Reasoning 20

  0.2

  170 175

  0.2

  165

  0.5

  160

  0.8

  155

  1

  [Kasabov 1996]

  Example Fuzzy Reasoning 2

‹

tables for fuzzy set definitions

  1 disapprove approve

  Approximate Reasoning 21 Advantages and Problems of Fuzzy Logic

  ‹

  inconsistency

  ‹

  inexact knowledge

  ‹

  inference

  ‹ inference mechanism ‹

  knowledge

  ‹

  linguistic variable

  membership function

  imprecision

  ‹

  non-monotonic reasoning

  ‹

  possibility

  ‹ probability ‹

  reasoning

  ‹

  rule

  ‹

  ‹

  ‹

  ‹ advantages

  

‹

fuzzy logic utilizes linguistic variables in combination with fuzzy rules and fuzzy inference in a formal approach to approximate reasoning

  ‹ foundation for a general theory of commonsense reasoning

  ‹ many practical applications ‹ natural use of vague and imprecise concepts

  ‹ hardware implementations for simpler tasks

  ‹ problems

  ‹ formulation of the task can be very tedious

  ‹ membership functions can be difficult to find ‹ multiple ways for combining evidence

  ‹ problems with long inference chains ‹ efficiency for complex tasks

  Approximate Reasoning 22 Summary Approximate Reasoning

  

‹

attempts to formalize some aspects of common- sense reasoning

  ‹ allows a more natural formulation of some types of

  ‹ fuzzy value ‹ fuzzy variable

  problems

  ‹ successfully applied to many real-world problems ‹ some fundamental and practical limitations

  ™ semantics, usage, efficiency Approximate Reasoning 23

  Important Concepts and Terms ‹ approximate reasoning

  ‹ common-sense reasoning ‹ crisp set

  ‹ default reasoning ‹ defuzzification

  ‹ extension principle

  ‹ fuzzification ‹ fuzzy inference

  ‹ fuzzy rule ‹ fuzzy set

  uncertainty