Weighted and Multi-Valued Automata: Comparison and Transfer of Ideas

Download as pdf or txt
Download as pdf or txt
You are on page 1of 25

Weighted and Multi-Valued Automata:

Comparison and Transfer of Ideas

Miroslav Ćirić
Department of Computer Science
Faculty of Sciences
University of Niš, Serbia
miroslav.ciric@pmf.edu.rs

Erasmus+ Compact Lecture on Weighted Automata and Networks


Institut für Informatik, Universität Leipzig, Germany
May 27 – June 23, 2018

1 Miroslav Ćirić Weighted and Multi-Valued Automata


Weighted and multi-valued automata

A = (A, σ, τ, {δx}x∈X )
⋆ A – set of states with |A| = n
⋆ X – fixed input alphabet
⋆ σ, τ : A → K – initial and final vectors with entries in K
⋆ δx : A × A → K – transition matrices with entries in K

⋆ weighted automata:
K – semiring, bimonoid . . . weights
⋆ multi-valued automata or fuzzy automata:
K – ordered algebraic structure truth values

2 Miroslav Ćirić Weighted and Multi-Valued Automata


Weights

⋆ model certain quantitative properties


⋆ amount of resources needed for the execution of a transition

⋆ time needed for the execution


⋆ cost of the execution
⋆ probability of successful execution of a transition
⋆ reliability of successful execution . . .

⋆ operations on weights – accumulation of weights


⋆ distributivity is not neccessary (bimonoids . . . )

3 Miroslav Ćirić Weighted and Multi-Valued Automata


Truth values

⋆ multi-valued logic – graded truth or intermediate truth


⋆ subtle nuances in reasoning – modeling of uncertainty
⋆ representation of imprecise aspects of human knowledge

⋆ ordering is essential – comparison of truth values


⋆ operations on truth values – logic conjectives
⋆ [0, 1] (real unit interval), max, min – Gödel structure
⋆ [0, 1], t-norm, t-conorm – Łukasiewicz and product structure

⋆ linearity of the ordering is not essential


⋆ truth values – lattices, ordered algebraic structures

4 Miroslav Ćirić Weighted and Multi-Valued Automata


Structures of truth values

Classical Multi-valued logics with Multi-valued logics with


Boolean logic linearly ordered more general structures
structures of truth values
of truth values (not necessarily linearly
ordered)
two-element structures on [0, 1] lattices, residuated
Boolean algebra determined by t-norms latices, etc.

5 Miroslav Ćirić Weighted and Multi-Valued Automata


Structures of truth values (cont.)

⋆ lattices: finite infimum and supremum –


conjunction and disjunction (intersection and union)
⋆ complete lattices: infinite infimum and supremum –
universal and existential quantifiers

⋆ infimum does not necessarily distribute over supremum


(except in distributive lattices)
⋆ new operation: multiplication ⊗ – distributes over suprema
strong conjunction

⋆ lattice ordered monoid


monoid + partial order (compatible w.r.t. multiplication ⊗)
lattice w.r.t. this partial order
⊗ distributes over (finite) suprema
(∨, ⊗)-reduct – semiring reduct
6 Miroslav Ćirić Weighted and Multi-Valued Automata
Residuation. Residuated functions

⋆ How to model the implication?

⋆ Residuated function
⋆ (P, 6), (Q, 6) – partially ordered sets, f : P → Q
⋆ f is residuated if there is g : Q → P satisfying

f (x) 6 y ⇔ x 6 g(y)

⋆ residuation property
⋆ if exists, such g is unique
⋆ it is called the residual of f and denoted by f ♯

7 Miroslav Ćirić Weighted and Multi-Valued Automata


Theorem on residuated functions

Theorem on residuated functions


The following conditions for f : P → Q are equivalent:
(i) f is residuated;
(ii) f is isotone and there is an isotone g : Q → P such that
IP 6 f ◦ g, g ◦ f 6 IQ ;
(iii) the inverse image under f of every principal down-set of Q is a
principal down-set of P;
(iv) f is isotone and the set {x ∈ P | f (x) 6 y} has the greatest element,
for every y ∈ Q.

⋆ principal down-set: a ↓= {x ∈ P | x 6 a}
⋆ lattice-theoretical counterpart of a continuous function
⋆ f ♯ (y) = ⊤{x ∈ P | f (x) 6 y} (⊤H – greatest element of H)
8 Miroslav Ćirić Weighted and Multi-Valued Automata
Residuated algebraic structures. Residuated semigroups

⋆ (S, ⊗, 6) – ordered semigroup


⋆ 6 is compatible w.r.t. ⊗
⋆ for a ∈ S, functions λa , ̺a : S → S are defined by
λa (x) = a ⊗ x, ̺a (x) = x ⊗ a
⋆ λa – inner left translation w.r.t. a
̺a – inner right translation w.r.t. a
⋆ residuated semigroup – λa and ̺a are residuated functions
⋆ a\b = λ♯a (b) = ⊤{x ∈ S | a ⊗ x 6 b} – right residual of b by a
⋆ b/a = ̺a♯ (b) = ⊤{x ∈ S | x ⊗ a 6 b} – left residual of b by a
⋆ residuation property
a ⊗ b 6 c ⇔ b 6 a\c ⇔ a 6 c/b

9 Miroslav Ćirić Weighted and Multi-Valued Automata


Quantales

⋆ Quantale
ordered semigroup (S, ⊗, 6)
complete lattice w.r.t. 6
⊗ distributes over arbitrary suprema (finite and infinite)

⋆ ⊗ is not necessarily commutative


⋆ inifinite distributivity ⇒ existence of residuals
_
a\b = {x ∈ S | a ⊗ x 6 b} = ⊤{x ∈ S | a ⊗ x 6 b}
_
b/a = {x ∈ S | x ⊗ a 6 b} = ⊤{x ∈ S | x ⊗ a 6 b}
⋆ unital quantale – with a multiplicative unit e
⋆ (S, ∨, ⊗, 0, e) – semiring (semiring reduct)
⋆ integral quantale – e is the greatest element (e = 1)

10 Miroslav Ćirić Weighted and Multi-Valued Automata


Residuated lattices

⋆ general meaning: lattice-ordered residuated semigroup


⋆ not necessarily commutative (left and right residuals)
⋆ not necessarily complete, not necessarily bounded
⋆ multi-valued logic – requires commutativity and completeness

⋆ Residuated lattice – algebra L = (L, ∨, ∧, ⊗, →, 0, 1)


(L, ∨, ∧, 0, 1) – bounded lattice with the least element 0
and the greatest element 1
(L, ⊗, 1) – commutative monoid with the unit 1
⊗ and → satisfy the residuation property
x⊗y6z ⇔ x6y→z

⋆ Complete residuated lattice – the lattice reduct is complete


commutative integral quantale
11 Miroslav Ćirić Weighted and Multi-Valued Automata
Residuated lattices (cont.)

⋆ only one residual (left and right residuals coincide)


⋆ operation → : residuum or residual implication
models the implication
⋆ residuation property x ⊗ y 6 z ⇔ x 6 y → z
modus ponens rule
deduction theorem
⋆ biresiduum or residual equivalence:
x ↔ y = (x → y) ∧ (y → x)
models the equivalence
⋆ negation: ¬x = x → 0

12 Miroslav Ćirić Weighted and Multi-Valued Automata


Special residuated structures

on [0, 1] with x ∧ y = min(x, y) and x ∨ y = max(x, y)


⋆ Göedel structure
(
1 if x 6 y
x ⊗ y = min(x, y), x→y=
y otherwise
⋆ Product structure or Goguen structure
(
1 if x 6 y
x ⊗ y = x · y, x→y= y
x
otherwise
⋆ Łukasiewicz structure
x ⊗ y = max(x + y − 1, 0), x → y = min(1 − x + y, 1)

⋆ Göedel and Łukasiewicz structure on finite chains in [0, 1]


⋆ Heyting algebra: L with ⊗ = ∧ (bounded Brouwer lattice)

13 Miroslav Ćirić Weighted and Multi-Valued Automata


Multi-valued structures and logics
Quantales
Substructural logics

Complete residuated Complete


lattices orthomodular lattices
Residuated logics Quantum logic

MTL-algebras
MTL-logic
Standard MTL-algebras

BL-algebras
Heyting algebras
Basic logic
Intuitionistic logic
Standard BL-algebras

Product algebras MV-algebras Gödel algebras


(Goguen algebras) (Wajsberg algebras)
Product logic Łukasiewicz logic Gödel logic
Product structure Łukasiewicz structure Gödel structure

Boolean algebras
Classical logic
Boolean structure

14 Miroslav Ćirić Weighted and Multi-Valued Automata


Residuation: some historical notes

⋆ Dedekind (1894) – quantales of ideals of rings


⋆ Schröder, Algebra und Logik der Relative (Leipzig, 1895)
quantales of binary relations
⋆ Brouwer (1920s) – relative pseudo complementation
⋆ Heyting (1930) – Heyting algebras
⋆ Ward, Dilworth (1930s) – (noncommutative) residuated
lattices, arithmetical applications
⋆ Mulvey (1986) – quantale of closed linear subspaces of a
non-commutative C⋆ -algebra
applications in functional analysis, topology
Gelfand, von Neumann, and Hilbert quantales

15 Miroslav Ćirić Weighted and Multi-Valued Automata


State reduction

⋆ Mimimization of NFA – computationally hard problem


⋆ the same goes for weighted and multi-valued automata
⋆ more practical – state reduction problem
give an efficient construction of a reasonably small automaton
(not necessarily minimal) equivalent to the given automaton

⋆ How to make a state reduction?

⋆ our main ideas came from algebra – quotient algebra


⋆ congruences – compatible equivalence relations
⋆ elements of the quotient algebra – equivalence classes
⋆ rows or columns in the correspodning Boolean matrix

16 Miroslav Ćirić Weighted and Multi-Valued Automata


Row automata

⋆ (S, +, ·, 0, e) – semiring with the unit e


⋆ A = (A, σ, τ, {δx}x∈X ) – weighted finite automaton over S
⋆ π ∈ SA×A – given matrix

⋆ Our idea: Construct an WFA whose states would be rows of π

⋆ aπ – a-row of π, πb – b-column of π
⋆ A – the set of all different rows of π
⋆ define σ, τ : A → S, δx : A × A → S, x ∈ X by

σ(aπ) = (σ · π)(a) = σ · (πa)


τ(aπ) = (π · τ)(a) = (aπ) · τ
δx (aπ, bπ) = (π · δx · π)(a, b) = (aπ) · δx · (πb)

17 Miroslav Ćirić Weighted and Multi-Valued Automata


Quasi-order matrices

⋆ Question: Are these definitions good?


⋆ Answer: Not necessarily.
If aπ = a′ π and bπ = b′ π, it does not have to be
σ · (πa) = σ · (πa′ ) or (aπ) · δx · (πb) = (a′ π) · δx · (πb′ )

⋆ Question: Under what conditions the definitions are good?

⋆ we need a partial order 6 on S (not necessarily compatible)


⋆ a square matrix π : A × A → S is
reflexive if e 6 π(a, a), for all a ∈ A
transitive if π(a, b) · π(b, c) 6 π(a, c), for all a, b, c ∈ A

⋆ Quasi-order matrix – reflexive and transitive matrix

18 Miroslav Ćirić Weighted and Multi-Valued Automata


Quasi-order matrices (cont.)

Theorem 1
Let π ∈ SA×A be a quasi-order matrix and a, b ∈ A. Then the following
conditions are equivalent:
(i) π(a, b) = π(b, a) = e
(ii) aπ = bπ
(iii) πa = πb

Theorem 2
Let π ∈ SA×A be a quasi-order matrix.
Then σ, τ and δx are well-defined and A = (A, σ, τ, {δx }x∈X ) is an WFA
satisfying |A| 6 |A|.

A – row automaton – isomorphic to column automaton

19 Miroslav Ćirić Weighted and Multi-Valued Automata


Notes on quasi-order matrices

⋆ S – semiring with unit e, 6 – partial order on S


⋆ partial order on SA×A is defined entrywise
µ 6 η ⇔ µ(a, b) 6 η(a, b), for all a, b ∈ S
⋆ if the ordering on S is compatible, then the ordering of matrices is
also compatible
⋆ for a matrix π with entries in a lattice-ordered monoid
π is reflexive ⇔ ∆ 6 π ⇒ π 6 π2
π is transitive ⇔ π2 6 π
π is a quasi-order matrix ⇒ π2 = π
⋆ ∆(a, a) = e (the unit), ∆(a, b) = 0, for a , b – unit matrix
⋆ Question: Under what conditions this holds for matrices with
entries in a semiring?
20 Miroslav Ćirić Weighted and Multi-Valued Automata
Notes on quasi-order matrices (cont.)

⋆ S – positively ordered semiring – compatible partial order 6


and 0 is the least element
π is reflexive ⇒ ∆ 6 π ⇒ π 6 π2

⋆ to prove
π is transitive ⇒ π2 6 π
we need something like
a1 , . . . , as 6 a ⇒ a1 + . . . + as 6 a
the addition behaves somehow like supremum

⋆ Question: In which class of semirings all of this is true?

⋆ Answer: Additively idempotent semirings

21 Miroslav Ćirić Weighted and Multi-Valued Automata


Additively idempotent semirings

⋆ a + a = a, for every a ∈ S (equivalently 1 + 1 = 1)


⋆ positively partially ordered
⋆ partial ordering: a 6 b ⇔ a + b = b
⋆ supremum coincides with addition
⋆ every quasi-order matrix π satisfies π2 = π

⋆ Question: Why π2 = π is so important?

⋆ behaviour of the row automaton A (the general case)


[[A]](ε) = σ · π2 · τ
[[A]](x1 x2 · · · xk ) = σ · π2 · δx1 · π2 · δx2 · π2 · . . . · π2 · δxk · π2 · τ

⋆ with π2 = π we avoid squares

22 Miroslav Ćirić Weighted and Multi-Valued Automata


Equivalence of A and A (additively idempotent case)

⋆ Question: Is A equivalent to A? Answer: Not necessarily.

⋆ Question: Under what conditions they are equivalent?

[[A]](ε) = σ · τ
[[A]](x1 x2 · · · xk ) = σ · δx1 · δx2 · . . . · δxk · τ

[[A]](ε) = σ · π · τ
[[A]](x1 x2 · · · xk ) = σ · π · δx1 · π · δx2 · π · . . . · π · δxk · π · τ

⋆ π has to be a solution of the general system


σ·τ=σ·π·τ
σ · δx1 · δx2 · . . . · δxk · τ = σ · π · δx1 · π · δx2 · π · . . . · π · δxk · π · τ

23 Miroslav Ćirić Weighted and Multi-Valued Automata


Notes on the general system

⋆ it may consist of infinitely many equations


⋆ can not be solved efficiently
⋆ we have to find as possible greater solutions (greater
solutions provide better reductions)
⋆ in the general case, there is no the greatest solution

⋆ instances of the general system


⋆ systems whose any solution is a solution to the general
system
⋆ we need instances with finitely many equations or inequations
which have the greatest solution and can be solved efficiently

to be continued . . .

24 Miroslav Ćirić Weighted and Multi-Valued Automata


In the next episodes

⋆ weakly linear systems

⋆ solving weakly linear systems

⋆ weakly linear systems in the state reduction

⋆ weakly linear systems in the study of


simulations and bisimulations

25 Miroslav Ćirić Weighted and Multi-Valued Automata

You might also like