Download CF(R1), CF(R2)

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
Kunstig intelligens (IT-2702) - høst 2006.
Forelesning 7
Emner:
Usikkerhetsbehandling
- Utvidelse av standard logikk
- Den klassiske ”certainty factor” metoden
- Fuzzy mengder
- Statistikk-baserte metoder
- og spesielt Bayesianske nett
- Kunnskapsbaserte metoder
Den reelle verden er usikker
It is the mark of an instructed mind to rest satisfied with that degree of precision
which the nature of the subject admits, and not to seek exactness where only
an approximation of the truth is possible.
(Aristotle)
All traditional logic habitually assumes that precise symbols are being employed.
It is therefore not applicable to this terrestrial life but only to an imagined
celestial existence.
(Bertrand Russell)
So far as the laws of mathematics refer ro reality they are not certain. And so far
as they are certain they do not refer to reality.
(Albert Einstein)
Inferens-metoder
• Deduksjon
- sannhetsbevarende slutning
- basis er slutningsregelen modus ponens
(P(x) -> Q(x)) & P(a) -> Q(a)
Klassisk eksempel:
&
->
(Isa-man(x) -> Is-mortal(x))
Isa-man(Socrates)
Is-mortal(Socrates)
• Abduksjon
- ikke sannhetsbevarende
- "inference to the best explanation"
(P(x) -> Q(x)) & Q(a) ~> P(a)
Eksempel:
(Has-appendicitis(x) -> Has-abdominal-pain(x))
& Has-abdominal-pain(Socrates)
~> Has-appendicitis(Socrates)
Ikke-monotone systemer
Forutsetninger for 1. ordens predikatlogikk:
- komplett domenebeskrivelse
- konsistent domenebeskrivelse
- monotont voksende kunnskapsbase
I ikke-monotone systemer er en eller flere av disse forutsetningene
ikke oppfylt.
Logikk-tilnærminger:
Modal-operatorer
Truth Maintenance systemer
unless, is-consistent-with, …
JTMS, ATMS, . . .
Andre
CWA, Circumscription, . . .
Modal-operatorer:
p(X) unless q(X) => r(X)
good-student(X) ^ M study-hard(X) => graduates(X)
is-consistent-with
Set cover approach:
En abduktiv forklaring på et sett av fakta (S2) er et annet sett av fakta
(S1) som er tilstrekkelig for å forårsake S2.
En optimal forklaring er det minimale sett S1.
Logikkbasert approach:
En abduktiv forklaring på et sett av observasjoner (O) er det minimale sett av
hypoteser (H) som er konsistent med den aktuelle bakgrunnskunnskap (K).
O kan ikke være utledbar fra K alene.
Usikkerhetsbehandling - Certainy Factors
- usikkerhetsanslag i regel-baserte systemer
- benyttet i MYCIN og avledede ES-skall
- basert på anslag av
- degree of belief
- degree of disbelief
MB(H/E)
MD(H/E)
- som kobineres i en Certainty Factor
CF(H/E) = MB(H/E) - MD(H/E)
Usikkerhetsbehandling - CF (forts.)
- eksempel, Mycin-type regel:
IF
(P1 and P2) or P3
THEN R1 (0.7) and R2 (0.3)
- kombinasjon av to regler som peker på samme konklusjon:
CF(R1)+CF(R2) - (CF(R1)xCF(R2))
|
CF(R1), CF(R2) pos.
CF(R1)+CF(R2) + (CF(R1)xCF(R2)) |
CF(R1), CF(R2) neg.
CF(R1)+CF(R2)
|
1 - min((abs(CF(R1), abs(CF(R2))
ellers
Usikkerhetsbehandling - Fuzzy Sets
- et fuzzy set (fose mendge?) er en mendge der elementene
i større eller mindre grad kan sies å være medlem av
mendgen
- en medlemsskapsfunksjon definerer i hvilken grad
(mellom 0 og 1) et element er medlem av mengden
- øvelse: tegn medlemsskapsfunksjonene [0,1] for
ung og gammel i mendgen av aldre [1,100].
1
0.5
0
1
10
20
30
40
50
60
70
80
90
100
Fig 9.6 the fuzzy set representation for “small integers.”
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
7
Fig 9.7 A fuzzy set representation for the sets short, medium, and tall
males.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
8
Fig 9.8 The inverted pendulum and the angle θ and dθ/dt input values.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
9
Fig 9.9 The fuzzy regions for the input values θ (a) and dθ/dt (b).
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
10
Fig 9.10 The fuzzy regions of the output value u, indicating the movement
of the pendulum base.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
11
Fig 9.11 The fuzzificzation of the input measures X1 = 1, X2 = -4
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
12
Fig 9.10 The Fuzzy Associative Matrix (FAM) for the pendulum problem.
The input values are on the left and top.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
13
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
14
Fig 9.13 The fuzzy consequents (a) and their union (b). The centroid of the
union (-2) is the crisp output.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
15
Usikkerhetsbehandling - Statistiske metoder
Basisbegreper:
Prior probability
(a priori sanns., ubetinget sanns.)
Sannsynligheten, P, for en hendelse, A, uten at
noe informasjon er gitt:
P(A)
Posterior probability (a posteriori sanns., betinget sanns.)
Sannsynligheten, P, for en hendelse, A, gitt
informsjonen E:
P(A/E)
Kombinasjon av uavhengige (ubetingede) sanns.
P(A & B) = P(A) x P(B)
Probability theory, the general form of Bayes’ theorem
The application of Bayes’ rule to the car purchase problem:
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
23
Naïve Bayes, or the Bayes classifier, that uses the partition
assumption, even when it is not justified:
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
24
Fig 5.4 The Bayesian representation of the traffic problem with potential
explanations.
Table 5.4 The joint probability distribution for the traffic and construction
variables of Fig 5.3.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
25
Fig 9.14 The graphical model for the traffic problem, first introduced in
Section 5.3.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
19
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
20
= Belief Networks
Defining the d-separation of nodes in a belief network (after Pearl 1988)
Another way to express the same thing:
Nodes are independent of their non-descendants given their parents.
(d-separated)
E
R
(d-connected)
If Known
A
B
A
C
If Known
E
B
If Known
A
If Known
C
OR
Fig 9.15 a is a serial connection of nodes where influence runs between A
and B unless V is instantiated. 9.15b is a diverging connection,
where influence runs between V’s children, unless V is
instantiated. In 9.15c, a converging connection, if nothing is
known about V the its parents are independent, otherwise
correlations exist between its parents.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
21
Rød kule: Verdien er kjent (gjelder for de tre siste eksempeluttrykkene)
Fig 9.16 An example of a Bayesian probabilistic network, where the
probability dependencies are located next to each node. This
example is from Pearl (1988).
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
24
Table 9.4 The probability distribution for p(WS), a function of p(W) and
p(R) given the effect of S. We calculate the effect for x, where
R = t and W = t.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
23
A junction tree algorithm.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
25
Fig 9.17 A junction tree (a) for the Bayesian probabilistic network of (b).
Note that we started to construct the transition table for the
rectangle R, W.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
26
Fig 9.18 A Markov state machine or Markov chain with four states, s1, ..., s4
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
27
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
28
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
29
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
30
Fig 9.19 A hidden Markov model of two states designed for the coin
flipping problem. The aij values are determined by the elements
of the 2 x 2 transition matrix.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
31
Fig 9.20 A hidden Markov model for the coin flipping problem. Each coin
will have its own individual bias.
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
32
Fig 9.21 A PFSM representing a set of phonemically related English words.
The probability of each word occurring is below that word.
Adapted from Jurasky and Martin (2000).
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
33
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
34
Fig 9.22 A trace of the Viterbi algorithm on the paths of Fig 9.21. Rows
report the maximum value for Viterbi on each word for each input
value (top row). Adapted from Jurafsky and Martin (2000).
Luger: Artificial Intelligence, 5th edition. © Pearson Education Limited, 2005
35
Usikkerhetsbehandling - Forklaringsbasert
- modell-basert tilnærming
- kausale relasjoner oftest benyttet, men også andre relasjoner
- relasjonene i modellen antas usikre, og kan ha ”degree of belief” anslag
- usikkerhet begrenses ved multiple forklaringer,
dvs. en hypotese støttes i større eller mindre
grad av forklaringene som genereres i modellen
eks. ABEL (Stanford), HeartFailureModel (MIT),
”Endorsement Theory” (UMass), CREEK (NTNU-IDI)
Related documents