Dr. Matthew Ikl

1
Probabilistic Quantifier Logic for General
Intelligence An Indefinite Probabilities
Approach
Dr. Matthew Iklé Department of Mathematics and
Computer Science Adams State College
2
Probabilistic Logic Networks

• A Probabilistic Logic Inference System
• unifies probability and logic
• key component of the Novamente Cognition
  Engine, an integrative AGI system
• BUT independent of Novamente -- designed
  with ability to be incorporated into other
  systems

3
Probabilistic Logic Networks

• supports the full scope of inferences required
  within an intelligent system, including e.g.
  first and higher order logic, intensional and
  extensional reasoning, and so forth
• lends itself naturally to methods of
  inference control that are computationally
  tractable, and able to make use of the inputs
  provided by
• non-logical cognitive mechanisms

4

Weight of Evidence

• What is it?
• Why is it important?
• E.g. belief revision
• One approach weight of ev. interval
  width
• .2,.8 means less evidence than .4,.6
• Pei Wangs NARS system
• Walleys Imprecise Probabilities
• Heuristic approaches

5
Indefinite Probabilities

• primary measure of uncertainty utilized within
  PLN
• hybrid of Walleys Imprecise Probabilities
  and Bayesian credible intervals
• provides a natural mechanism for
  determining weight of evidence
• PLNs logical inference rules are associated
  with indefinite truth value formulas or
  procedures
• Prior papers have given indefinite truth
  value formulas for a number of PLN inference
  rules, but have not dealt with quantifiers

6
Indefinite Probabilities Review

• truth-value takes the form of a quadruple
  (L, U, b, k)
• There is a probability b that, after k more
  observations, the truth value assigned to the
  statement S will lie in the interval L, U
• Given intervals, Li,Ui , of mean premise
  probabilities, we first find a distribution
  from the second-order distribution family
  supported on
• L1i,U1i so that these means have L i,Ui as
  (100bi) credible intervals
• For each premise, we use Monte-Carlo
  methods to generate samples for each of
  the first-order distributions with means given
  by samples of the second- order distributions.
  We then apply the inference rules to the set of
  premises for each sample point, and calculate
  the mean of each of these distributions.

7
Quantifiers Via Indefinite Probabilities

• Goals
• logical and conceptual consistency
• agreement with standard quantifier logic for
  the crisp case (for all expressions to which
  standard quantifier logic assigns truth values)
• gives intuitively reasonable answers in
  practical cases
• compatibility with probability theory in
  general and PLN in particular
• handles fuzzy quantifiers as well as
  standard universal and existential quantifiers
• comprehensive, conceptually coherent,
  probabilistically grounded and computationally
  tractable approach

8
Quantifiers in Indefinite Probabilities

• utilize third-order probabilities
• non-standard semantic approach
• we assign truth values to expressions
  with unbound variables, yet without in doing so
  binding the variables
• unusual but not contradictory
• expression with unbound variables,
  as a mathematical entity, may be mapped into
  a truth value without introducing any
  mathematical or conceptual inconsistency
• approach reduces to the standard
  crisp approach in terms of truth value
  assignation for all expressions for which the
  standard crisp approach assigns a truth value.
  
• our approach also assigns truth
  values to some expressions (formulas standard
  crisp approach assigns no truth value

9
Quantifiers in Indefinite Probabilities ForAll

• Suppose we have an indefinite probability
  for an expression F(t) with unbound variable t,
  summarizing an envelope E of probability
  distributions corresponding to F(t)
• How do derive from this an indefinite
  probability for the expression ForAll x, F(x)?
• we consider the envelope E to be part of a
  higher-level envelope E1, which is an envelope
  of envelopes
• given that we have observed E, what is the
  chance (according to E1) that the true envelope
  describing the world actually is almost entirely
  supported within 1-e, 1, where the latter
  interval is interpreted to constitute
  essentially 1

10
Quantifiers in Indefinite Probabilities ThereExis
ts

• For ThereExists x, F(x), what is the chance
  (according to E1) that the true envelope
  describing the world actually is not entirely
  supported within 0, e, where the latter
  interval is interpreted to constitute
  essentially zero

11
Quantifiers in Indefinite Probabilities The
proxy_confidence_level parameter

• By almost entirely (in ForAll case) we mean that
  the fraction contained is at least
  proxy_confidence_level (PCL)
• the interval PCL, 1 represents the fraction of
  bottom-level distributions completely contained
  in the interval 1-e, 1

12
Quantifiers in Indefinite Probabilities Fuzzy
Quantifiers

• indefinite probabilities provide a natural
• method for fuzzy quantifiers such as AlmostAll
  and Afew
• In analogy with the interval PCL, 1 we
• introduce the parameters lower_proxy_confiden
  ce (LPC) and
• upper_proxy_confidence (UPC)
• Letting LPC, UPC 0.9, 0.99, the
• interval could now naturally represent AlmostAll
• the same interval could represent AFew by
  setting LPC to a value such
• as 0.05 and UPC to, say, 0.1.

13
Summary

• incorporating a third level of
  distributions, as perturbations, into the
  indefinite probabilities framework allows for
  extension of indefinite probabilities to handle
  a sliding scale of fuzzy and crisp quantifiers
• computationally tractable