Probabilistic Acceptors for. Languages over Infinite Words. Christel Baier. Technical University Dresden joint work with Nathalie Bertrand and Marcus GröÃer.
Probabilistic Acceptors for Languages over Infinite Words Christel Baier Technical University Dresden joint work with Nathalie Bertrand and Marcus Gr¨oßer
1 / 148
ω –regular languages
pba-1
• languages over infinite words L ⊆ Σω • arise from ω -regular expressions α1β ω1 + ... + αn β ωn where αi , β i are regular expressions with ε ∈ β i
2 / 148
ω –regular languages
pba-1
• languages over infinite words L ⊆ Σω • arise from ω -regular expressions α1β ω1 + ... + αn β ωn where αi , β i are regular expressions with ε ∈ β i • recognized by many types of ω -automata, e.g., NBA (nondeterministic B¨ uchi automata) syntax as NFA, but with acceptance criterion “visit infinitely often a final state”
3 / 148
ω –regular languages
pba-1
• languages over infinite words L ⊆ Σω • arise from ω -regular expressions α1β ω1 + ... + αn β ωn • recognized by many types of ω -automata, e.g., NBA (nondeterministic B¨ uchi automata) • can be used for verifying linear time properties M) ∩ L (A A) = ∅ with graph “check whether L (M algorithm in the product M × A ” where M is the system model, A is an NBA for the bad behaviors 4 / 148
ω -automata
pba-2
NBA
5 / 148
ω -automata
pba-2
NBA
DBA
6 / 148
ω -automata
pba-2
NBA
(a + b)∗ aω DBA
NBA b
1 a
a
2 a
7 / 148
ω -automata
pba-2
♦ F ♦ Hi → ♦ Ki ) NSA/DSA (Streett automata) ( i ♦ ¬Ki ∧ ♦ Hi ) NRA/DRA (Rabin automata) (♦ i .. . (a + b)∗ aω NBA
DBA
NBA b
1 a
a
2 a
8 / 148
ω -automata
pba-2
♦ F ♦ Hi → ♦ Ki ) NSA/DSA (Streett automata) ( i ♦ ¬Ki ∧ ♦ Hi ) NRA/DRA (Rabin automata) (♦ i .. . (a + b)∗ aω NBA
DBA exp
NBA DSA/DRA/... [Safra ’88]
NBA b
1 a
a
2 a
9 / 148
ω -automata
pba-2
♦ F ♦ Hi → ♦ Ki ) NSA/DSA (Streett automata) ( i ♦ ¬Ki ∧ ♦ Hi ) NRA/DRA (Rabin automata) (♦ i .. . (a + b)∗ aω probabilistic B¨ uchi automata ? DBA NBA
exp
NBA DSA/DRA/... [Safra ’88]
NBA b
1 a
a
2 a
10 / 148
Overview pba-3
• • • •
definition of probabilistic B¨uchi automata (PBA) expressiveness of PBA efficiency of PBA other acceptance conditions (Streett/Rabin, alternative semantics) • composition operators on PBA • decision problems for PBA
11 / 148
Probabilistic B¨ uchi Automaton (PBA)
pba-4
PBA: syntax as for NBA, but all choices are resolved probabilistically
a,
1 3
PBA probabilistic choice
q a,
2 3
a
q a
NBA nondeterministic choice
12 / 148
Probabilistic B¨ uchi Automaton (PBA)
pba-4
P = (Q, δ , µ , F) over alphabet Σ : • Q finite state space • transition probability function δ : Q × Σ × Q → [0, 1]
13 / 148
Probabilistic B¨ uchi Automaton (PBA)
pba-4
P = (Q, δ , µ , F) over alphabet Σ : • Q finite state space • transition probability function δ : Q × Σ × Q → [0, 1] s.t. for all q ∈ Q, a ∈ Σ : δ (q, a, p) ∈ {0, 1} ∈Q p∈
14 / 148
Probabilistic B¨ uchi Automaton (PBA)
pba-4
P = (Q, δ , µ , F) over alphabet Σ : • Q finite state space • transition probability function δ : Q × Σ × Q → [0, 1] s.t. for all q ∈ Q, a ∈ Σ : δ (q, a, p) ∈ {0, 1} ∈Q p∈
• initial distribution µ : Q → [0, 1]
15 / 148
Probabilistic B¨ uchi Automaton (PBA)
pba-4
P = (Q, δ , µ , F) over alphabet Σ : • Q finite state space • transition probability function δ : Q × Σ × Q → [0, 1] s.t. for all q ∈ Q, a ∈ Σ : δ (q, a, p) ∈ {0, 1} ∈Q p∈
• initial distribution µ : Q → [0, 1] • set of final states F ⊆ Q
16 / 148
Semantics of PBA
pba-5
P = (Q, δ , µ , F) • Q finite state space • transition probability function δ : Q × Σ × Q → [0, 1] • initial distribution µ • set of final states F ⊆ Q
17 / 148
Semantics of PBA
pba-5
P = (Q, δ , µ , F) • Q finite state space • transition probability function δ : Q × Σ × Q → [0, 1] • initial distribution µ • set of final states F ⊆ Q For x ∈ Σω : Pr Pr(x) = probability for the accepting runs for x
18 / 148
Semantics of PBA
pba-5
P = (Q, δ , µ , F) • Q finite state space • transition probability function δ : Q × Σ × Q → [0, 1] • initial distribution µ • set of final states F ⊆ Q For x ∈ Σω : Pr Pr(x) = probability for the accepting runs for x accepting run: visits F infinitely often
19 / 148
Semantics of PBA P = (Q, δ , µ , F)
pba-5
←
can be viewed as a Markov decision process
• Q finite state space • transition probability function δ : Q × Σ × Q → [0, 1] • initial distribution µ • set of final states F ⊆ Q For x ∈ Σω : Pr Pr(x) = probability for the accepting runs for x probability measure in the infinite Markov chain induced by x viewed as a scheduler 20 / 148
Semantics of PBA P = (Q, δ , µ , F)
pba-5
←
can be viewed as a Markov decision process
• Q finite state space • transition probability function δ : Q × Σ × Q → [0, 1] • initial distribution µ • set of final states F ⊆ Q For x ∈ Σω : Pr Pr(x) = probability for the accepting runs for x P ) = x ∈ Σω : Pr accepted language: L(P Pr(x) > 0 21 / 148
Example for a PBA a, 12
1
pba-6
2
b,1 a,
1 2
a,1
final state
non-final state 22 / 148
Example for a PBA a, 12
1
pba-6
2
b,1 a,
1 2
a,1
input word x = aabaωω final state
non-final state 23 / 148
Example for a PBA a , 12
1
pba-6
1 1
2
b,1 a,
1 2
1 2 1 2
1 2
1
a,1
1
input word aabaω
x = aaba
1 1 .. .
1 2
1 2 1 2
.. .
1
a
2 b
1 2
1 2
a 2
2
1
1 2
a 2
1 2
2 .. .
1
a
1
a .. .
2 1
.. .
24 / 148
Example for a PBA a , 12
1
pba-6
1 1
2
b,1 a,
1 2
1
x = aaba
1 4
1 1 .. .
1 2
1 2 1 2
.. .
a 1
a
2 b
1 2
1 2
1 2
2 2
1
aabaω
1 1 Pr(x) = 18 + 16 + 32 + ...
1 2
1
a,1
input word
=
1 2 1 2
a 2
1 2
2 .. .
1
a
1
a .. .
2 1
.. .
25 / 148
Example for a PBA a , 12
1
pba-6
1 1
2
b,1 a,
1 2
1
x = aaba
1 1 .. .
1 2
1 2 1 2
.. .
a 1
a
2 reject
1 2
1 2
1 2
2 2
1
aabaω
1 1 Pr(x) = 18 + 16 + 32 + ...
1 2
1
a,1
input word
= 1-Pr(xx is rejected)
1 2 1 2
b a
2
1 2
2 .. .
1
a
1
a .. .
2 1
.. .
26 / 148
Example for PBA a,
1
1 2
pba-7
accepted language: 2
(a + b)∗aω
b,1 a,
1 2
a,1
27 / 148
Example for PBA a,
1
1 2
pba-7
accepted language: 2
(a + b)∗aω
b,1 a,
1 2
a,1
Thus: PBA are strictly more expressive than DBA
28 / 148
Example for PBA a,
1
pba-7
1 2
accepted language: 2
(a + b)∗aω
b,1 a,
1 2
a,1
Thus: PBA are strictly more expressive than DBA
b 2
c
1 a,
1 2
b a,
1 2
3
29 / 148
Example for PBA a,
1
pba-7
1 2
accepted language: 2
(a + b)∗aω
b,1 a,
1 2
a,1
Thus: PBA are strictly more expressive than DBA
b 2
c
1 a,
1 2
b a,
1 2
3
NBA accepts ((ac)∗ ab)ω 30 / 148
Example for PBA a,
1
pba-7
1 2
accepted language: 2
(a + b)∗aω
b,1 a,
1 2
a,1
Thus: PBA are strictly more expressive than DBA
b 2
c
1 a,
1 2
b a,
1 2
accepted language: (ab + ac)∗(ab)ω
3
but NBA accepts ((ac)∗ab)ω 31 / 148
Expressiveness of PBA
pba-8a
32 / 148
Expressiveness of PBA
pba-8
Theorem: PBA are strictly more expressive than NBA
33 / 148
PBA are strictly more expressive than NBA
pba-8a
from NBA to PBA: NBA
NBA deterministic in = PBA limit Courcoubetis/ Yannakakis ’95
34 / 148
PBA are strictly more expressive than NBA
pba-8a
from NBA to PBA: NBA
NBA deterministic in = PBA limit Courcoubetis/ Yannakakis ’95 d
d
a a b
c b 35 / 148
PBA are strictly more expressive than NBA
pba-8a
from NBA to PBA: NBA
NBA deterministic in = PBA limit Courcoubetis/ Yannakakis ’95 d
d
a a b
c b
deterministic 36 / 148
PBA are strictly more expressive than NBA
pba-8a
from NBA to PBA: NBA
NBA deterministic in = PBA limit Courcoubetis/ Yannakakis ’95 d, 12
d d
a a b
= c b
d, 12
a, 12 a, 12
c b b
deterministic 37 / 148
PBA are strictly more expressive than NBA
pba-8
• from NBA to PBA: via NBA that are deterministic in limit ω -regular languages • PBA can accept non-ω
38 / 148
PBA are strictly more expressive than NBA
pba-8
• from NBA to PBA: via NBA that are deterministic in limit ω -regular languages • PBA can accept non-ω b 1 a , 12
2 a,
1 2
a
39 / 148
PBA are strictly more expressive than NBA
pba-8
• from NBA to PBA: via NBA that are deterministic in limit ω -regular languages • PBA can accept non-ω b 1 a , 12
2 a,
1 2
a
accepted language: ak11bak2 bak3 b . . . :
40 / 148
PBA are strictly more expressive than NBA
pba-8
• from NBA to PBA: via NBA that are deterministic in limit ω -regular languages • PBA can accept non-ω b 1 a , 12
2 a,
1 2
a
accepted language: ∞
1 k i k11 k2 k3 a ba ba b . . . : 1− >0 2 i=1 i=1
41 / 148
PBA are strictly more expressive than NBA
pba-uniform
• from NBA to PBA: via NBA that are deterministic in limit ω -regular languages • PBA can accept non-ω b 1 a , 12
2 a,
1 2
a
• uniform PBA cover exactly the class of ω -regular languages
42 / 148
PBA are strictly more expressive than NBA
pba-uniform
• from NBA to PBA: via NBA that are deterministic in limit ω -regular languages • PBA can accept non-ω b 1 a , 12
2 a,
1 2
a
• uniform PBA cover exactly the class of ω -regular languages where “uniformity” is a probabilistic condition on end components 43 / 148
Efficiency of PBA
pba-12
44 / 148
Efficiency of PBA
pba-12
Theorem: There exist uniform PBA Pn with |Pn | = O(n) s.t. each equivalent NSA An has Ω(2n /n) states. NSA: nondeterministic Streett automaton
45 / 148
Efficiency of PBA
pba-12
Theorem: There exist uniform PBA Pn with |Pn | = O(n) s.t. each equivalent NSA An has Ω(2n /n) states. Proof: consider the ω -regular languages Ln = xyω : x, y ∈ {a, b}∗ , |y| = n
46 / 148
Efficiency of PBA
pba-12
Theorem: There exist uniform PBA Pn with |Pn | = O(n) s.t. each equivalent NSA An has Ω(2n /n) states. Proof: consider the ω -regular languages Ln = xyω : x, y ∈ {a, b}∗, |y| = n show that 1. Each any NSA for Ln has ≥ 2n /n states. 2. There is a uniform PBA with O(n) states. 47 / 148
Efficiency of PBA
pba-12
Theorem: There exist uniform PBA Pn with |Pn | = O(n) s.t. each equivalent NSA An has Ω(2n /n) states. Proof: consider the ω -regular languages Ln = xyω : x, y ∈ {a, b}∗ , |y| = n Each any NSA for Ln has ≥ 2n /n states
48 / 148
Efficiency of PBA
pba-12
Theorem: There exist uniform PBA Pn with |Pn | = O(n) s.t. each equivalent NSA An has Ω(2n /n) states. Proof: consider the ω -regular languages Ln = xyω : x, y ∈ {a, b}∗ , |y| = n Each any NSA for Ln has ≥ 2n /n states as the “accepting loops” for y1 ≡ y2 , |y1| = |y2| = n are disjoint where for y1 = c1 c2 . . . cn and y2 = d1 d2 . . . dn : y1 ≡ y2 iff ∃i s.t. c1 c2 . . . cn = di . . . dn d1 . . . di−1 49 / 148
Efficiency of PBA
pba-13
Thm: There exist uniform PBA Pn with |Pn | = O(n) s.t. each equivalent NSA An has Ω(2n /n) states. Proof: consider the ω -regular languages Ln = xyωω : x, y ∈ {a, b}∗ , |y| = n show that 1. each NSA for Ln has ≥ 2n /n states
√
2. there exists a uniform PBA with 2n states
50 / 148
Efficiency of PBA
pba-13
Thm: There exist uniform PBA Pn with |Pn | = O(n) s.t. each equivalent NSA An has Ω(2n /n) states. Proof: consider the ω -regular languages Ln = xyωω : x, y ∈ {a, b}∗ , |y| = n PBA for Ln mit 2n states: 1a
2a
3a
...
na
1b
2b
3b
...
nb 51 / 148
Efficiency of PBA
pba-13
Thm: There exist uniform PBA Pn with |Pn | = O(n) s.t. each equivalent NSA An has Ω(2n /n) states. Proof: consider the ω -regular languages Ln = xyωω : x, y ∈ {a, b}∗ , |y| = n PBA for Ln mit 2n states: a a a
1a
2a a
1b
a
a
3a
...
na
3b
...
nb
a 2b
a
X 52 / 148
Efficiency of PBA
pba-13
Thm: There exist uniform PBA Pn with |Pn | = O(n) s.t. each equivalent NSA An has Ω(2n /n) states. Proof: consider the ω -regular languages Ln = xyωω : x, y ∈ {a, b}∗ , |y| = n PBA for Ln mit 2n states: a a b b a 1a 2 3a a a a a a a 1b
2b
3b
a b ... a ...
b a
na
nb
a
X 53 / 148
Efficiency of PBA
pba-13
Thm: There exist uniform PBA Pn with |Pn | = O(n) s.t. each equivalent NSA An has Ω(2n /n) states. Proof: consider the ω -regular languages Ln = xyωω : x, y ∈ {a, b}∗ , |y| = n PBA for Ln mit 2n states: a a b b a 1a 2 3a a a a a a a 1b
2b
3b
a b ... a ...
b a
na b
X
a
X
nb
54 / 148
Efficiency of PBA
pba-13
Thm: There exist uniform PBA Pn with |Pn | = O(n) s.t. each equivalent NSA An has Ω(2n /n) states. Proof: consider the ω -regular languages Ln = xyωω : x, y ∈ {a, b}∗ , |y| = n PBA for Ln mit 2n states: a a b b a 1a 2 3a a a ba b aa a b b 3b b 1b a 2b a b
b
a b ... a
b a
na b
X
b ... a
b a
nb
a
X
b
55 / 148
uPBA for L3 = {xyω : x, y ∈ {a, b}∗ , |y| = 3 } a a b
b 1b
a
a a b
1a
a b b
pba-14
2a b a 2b
a b
3a
a b
3b
uniform distributions
b
56 / 148
uPBA for L3 = {xyω : x, y ∈ {a, b}∗ , |y| = 3 } a a b
b 1b
a
a a b
1a
a b b
pba-14
2a b a 2b
a b
3a
a b
3b
uniform distributions
b
1. Let z = c1c2 . . . ∈ {a, b}ω s.t. ∞
∃ i with ci = a and ci+3 = b
57 / 148
uPBA for L3 = {xyω : x, y ∈ {a, b}∗ , |y| = 3 } a a b
b 1b
a
a a b
1a
a b b
pba-14
2a b a 2b
a b
3a
a b
3b
uniform distributions
b
1. Let z = c1c2 . . . ∈ {a, b}ω s.t. ∞
∃ i with ci = a and ci+3 = b c11...cii−1 −1a
ci +1
ci +3
b
→ with prob. 1: ∃i s.t. −−−−−→ 1a −−→ 2a −−→= 3a − 58 / 148
uPBA for L3 = {xyω : x, y ∈ {a, b}∗ , |y| = 3 } a a b
b 1b
a
a a b
1a
a b b
pba-14
2a b a 2b
a b
3a
a b
3b
uniform distributions
b
1. Let z = c1c2 . . . ∈ {a, b}ω s.t. ∞
∃ i with ci = a and ci+3 = b c11...cii−1 −1a
ci +1
ci +3
b
→ with prob. 1: ∃i s.t. −−−−−→ 1a −−→ 2a −−→= 3a − i.e. z will be rejected almost surely
59 / 148
uPBA for L3 = {xyω : x, y ∈ {a, b}∗ , |y| = 3 }
a a b
b 1b
a
a a b
1a
a b b
pba-15
2a
a b
3a
2b
a b
3b
b a
uniform distributions
b
2. Let z = xyω ∈ L3 where x = c1 . . . ci .
60 / 148
uPBA for L3 = {xyω : x, y ∈ {a, b}∗ , |y| = 3 }
a a b
b 1b
a
a a b
1a
a b b
pba-15
2a
a b
3a
2b
a b
3b
b a
uniform distributions
b
2. Let z = xyω ∈ L3 where x = c1 . . . ci . All runs for z that start with the prefix c22 ci c1 1a − → 1c11 − → ... − → 1ci are infinite 61 / 148
uPBA for L3 = {xyω : x, y ∈ {a, b}∗ , |y| = 3 }
a a b
b 1b
a
a a b
1a
a b b
pba-15
2a
a b
3a
2b
a b
3b
b a
uniform distributions
b
2. Let z = xyω ∈ L3 where x = c1 . . . ci . All runs for z that start with the prefix c22 ci c1 1a − → 1c11 − → ... − → 1ci are infinite (and accepting). 62 / 148
uPBA for L3 = {xyω : x, y ∈ {a, b}∗ , |y| = 3 }
a a b
b 1b
a
a a b
1a
a b b
pba-15
2a
a b
3a
2b
a b
3b
b a
uniform distributions
b
2. Let z = xyω ∈ L3 where x = c1 . . . ci . All runs for z that start with the prefix c22 ci c1 1a − → 1c11 − → ... − → 1ci are infinite (and accepting). 1 i Hence: Pr(z) ≥ ( 2 ) > 0 63 / 148
Efficiency of PBA
pba-16
Thm: There exist uniform PBA with of size O(n) s.t. each equivalent NSA has Ω(2n /n) states Thm: There exist NBA A n of size O(n) s.t. each equivalent uniform PBA has Ω(2n ) states
64 / 148
Efficiency of PBA
pba-16
Thm: There exist NBA A n of size O(n) s.t. each equivalent uniform PBA has Ω(2n ) states Proof: consider the language ((a + b)∗a(a + b)n−1c)ω
65 / 148
Efficiency of PBA
pba-16
Thm: There exist NBA A n of size O(n) s.t. each equivalent uniform PBA has Ω(2n ) states Proof: consider the language ((a + b)∗a(a + b)n−1c)ω NBA: 0 a, b
a
1
a, b
2
a, b . . .
a, b
n−1 c
a, b
n
66 / 148
Efficiency of PBA
pba-16
Thm: There exist NBA A n of size O(n) s.t. each equivalent uniform PBA has Ω(2n ) states Proof: consider the language ((a + b)∗a(a + b)n−1c)ω NBA: 0 a, b
a
1
a, b
2
a, b . . .
a, b
n−1 c
a, b
n
PBA: has to “remember” all a ’s of the last n symbols
67 / 148
Probabilistic ω -automata
pba-17
accepted language: P ) = x ∈ Σ ω : PrP ( accepting runs for x )} > 0 L (P acceptance conditions: • B¨uchi: ♦ F ♦Hi ∧ ♦¬Ki ) • Rabin: ( 1≤i≤k ♦ Hi → ♦ Ki ) ( • Streett: 1≤i≤k
68 / 148
Probabilistic ω -automata
pba-17
accepted language: P ) = x ∈ Σ ω : PrP ( accepting runs for x )} > 0 L (P acceptance conditions: • B¨uchi: ♦ F ♦Hi ∧ ♦¬Ki ) • Rabin: ( 1≤i≤k ♦ Hi → ♦ Ki ) ( • Streett: 1≤i≤k
poly
PRA −−→ PBA as for nondeterministic automata exp PSA −→ PBA preserving uniformity 69 / 148
Probabilistic ω -automata
pba-17
accepted language: P ) = x ∈ Σ ω : PrP ( accepting runs for x )} > 0 L (P acceptance conditions: • B¨uchi: ♦ F ♦Hi ∧ ♦¬Ki ) • Rabin: ( 1≤i≤k ♦ Hi → ♦ Ki ) ( • Streett: 1≤i≤k
poly
PRA −−→ PBA as for nondeterministic automata exp PSA −→ PBA preserving uniformity poly
PSA −−→ PBA possibly non-uniform 70 / 148
Poly transformation from PSA to PBA PSA P with acc. condition
pba-18
♦ Hi → ♦ Ki ) (
1≤i≤m
PBA consists of several copies of P
71 / 148
Poly transformation from PSA to PBA PSA P with acc. condition
pba-18
♦ Hi → ♦ Ki ) (
1≤i≤m
PBA: P init P accept
72 / 148
Poly transformation from PSA to PBA PSA P with acc. condition
pba-18
♦ Hi → ♦ Ki ) (
1≤i≤m
PBA: P init P accept Ki Pi
Hi
73 / 148
Poly transformation from PSA to PBA PSA P with acc. condition
pba-18
♦ Hi → ♦ Ki ) (
1≤i≤m
PBA: P init P accept Ki Pi
Hi Hj
74 / 148
Poly transformation from PSA to PBA PSA P with acc. condition
pba-18
♦ Hi → ♦ Ki ) (
1≤i≤m
PBA: P init P accept Ki Pi P i,j i = j
Hi Hj
Ki Hj 75 / 148
Poly transformation from PSA to PBA PSA P with acc. condition
pba-18
♦ Hi → ♦ Ki ) (
1≤i≤m
PBA: P init P accept Ki Pi P i,j i = j
Hi Hj
Pj
P j,k j = k
Ki Hj
76 / 148
Poly transformation from PSA to PBA PSA P with acc. condition
pba-18
♦ Hi → ♦ Ki ) (
1≤i≤m
PBA: P init
Ki Pi P i,j i = j
Hi Hj
Hj
P accept Kj Hi
P j,k j = k
Ki Hj
Hi
Pj
Kj 77 / 148
Poly transformation from PSA to PBA PSA P with acc. condition
pba-18
♦ Hi → ♦ Ki ) (
1≤i≤m
PBA: P init
Ki Pi P i,j i = j
Hi Hj
Hj
P accept
size: O(m2|P|) Kj Hi
P j,k j = k
Ki Hj
Hi
Pj
Kj 78 / 148
Alternative semantics for PBA
pba-19
standard semantics: P ) = x ∈ Σω : PrP (xx) > 0 L (P
79 / 148
Alternative semantics for PBA
pba-19
standard semantics: P ) = x ∈ Σω : PrP (xx) > 0 L (P alternative semantics: • threshold semantics • almost-sure semantics
80 / 148
Threshold semantics for PBA
pba-19
for PBA P and λ ∈]0, 1[ P ) = x ∈Σ Σωω : PrP (xx) > λ L>λ (P
81 / 148
Threshold semantics for PBA
pba-19
for PBA P and λ ∈]0, 1[ P ) = x ∈Σ Σωω : PrP (xx) > λ L>λ (P Results: P ) = L (P P) • ∀ PBA P ∀λ ∈]0, 1[ ∃ PBA P s.t. L>λ (P
82 / 148
Threshold semantics for PBA
pba-19
for PBA P and λ ∈]0, 1[ L>λ (P P ) = x ∈Σ Σωω : PrP (xx) > λ Results: P ) = L (P P) • ∀ PBA P ∀λ ∈]0, 1[ ∃ PBA P s.t. L>λ (P
P
ε, λ ε, 1 − λ
accept
a∈Σ
P
P is uniform if P is uniform 83 / 148
Threshold semantics for PBA
pba-19
for PBA P and λ ∈]0, 1[ P ) = x ∈Σ Σωω : PrP (xx) > λ L>λ (P Results: P ) = L (P P) • ∀ PBA P ∀λ ∈]0, 1[ ∃ PBA P s.t. L>λ (P P ) cannot be • ∃ PBA P ∃λ ∈]0, 1[ s.t. L>λ (P recognized by a standard PBA
84 / 148
Threshold semantics for PBA
pba-19
for PBA P and λ ∈]0, 1[ P ) = x ∈Σ Σωω : PrP (xx) > λ L>λ (P Results: P ) = L (P P) • ∀ PBA P ∀λ ∈]0, 1[ ∃ PBA P s.t. L>λ (P P ) cannot be • ∃ PBA P ∃λ ∈]0, 1[ s.t. L>λ (P recognized by a standard PBA P ) is not • ∃ uniform PBA P ∃λ ∈]0, 1[ s.t L>λ (P ω -regular
85 / 148
Almost-sure semantics of PBA P ) = x ∈ Σω : Pr(x) = 1 L=1 (P
pba-20a
86 / 148
Almost-sure semantics of PBA P ) = x ∈ Σω : Pr(x) = 1 L=1 (P
pba-20a
Theorem: For each PBA P there exists a PBA P P ) = L (P P ). such that L=1 (P
87 / 148
Almost-sure semantics of PBA P ) = x ∈ Σω : Pr(x) = 1 L=1 (P
pba-20a
Theorem: For each PBA P there exists a PBA P P ) = L (P P ). such that L=1 (P Proof sketch. PBA P for input word x • simulates P with input x • guesses at random a word position i • checks whether ¬F holds with positive probability from position i • if so, P rejects; otherwise P accepts. 88 / 148
Almost-sure semantics of PBA P ) = x ∈ Σω : Pr(x) = 1 L=1 (P
pba-20a
Theorem: For each PBA P there exists a PBA P P ) = L (P P ). such that L=1 (P P ) is Theorem: There exists a PBA P such that L (P not recognizable by a PBA with the almost-sure semantics.
89 / 148
Almost-sure semantics of PBA P ) = x ∈ Σω : Pr(x) = 1 L=1 (P
pba-20a
Theorem: For each PBA P there exists a PBA P P ) = L (P P ). such that L=1 (P P ) is Theorem: There exists a PBA P such that L (P not recognizable by a PBA with the almost-sure semantics. example: PBA for the ω-regular language (a + b)∗aω
90 / 148
PBA with standard and alternative semantics
pba-20
PBA
DBA
91 / 148
PBA with standard and alternative semantics
pba-20
PBA ω -regular languages uniform PBA DBA
92 / 148
PBA with standard and alternative semantics
pba-20
PBA with thresholds PBA ω -regular languages uniform PBA DBA
93 / 148
PBA with standard and alternative semantics
pba-20
PBA with thresholds PBA ω -regular languages uniform PBA DBA
PBA almost sure semantics
94 / 148
PBA with standard and alternative semantics
pba-20
PBA with thresholds (a+b)∗aω
PBA ω -regular languages uniform PBA DBA
PBA almost sure semantics
95 / 148
PBA with standard and alternative semantics
pba-20
PBA with thresholds (a+b)∗aω
PBA ω -regular languages uniform PBA DBA
{ak11bak2 b. . . :
∞
PBA almost sure semantics
(1 − ( 12 )kii ) > 0}
i=1
96 / 148
PBA with standard and alternative semantics
pba-20
PBA with thresholds (a+b)∗aω
PBA ω -regular languages uniform PBA PBA almost sure semantics
DBA
{ak11bak2 b. . . :
∞
(1 − ( 12 )kii ) > 0}
i=1
{ak11bak22b. . . :
∞
(1 − ( 12 )ki ) = 0}
97 / 148
PBA with standard and alternative semantics
pba-20
PBA with thresholds (a+b)∗aω
{ak11bak2 b. . . :
PBA, PSA, PRA ω -regular languages uniform PBA PBA almost sure semantics DBA ∞
(1 − ( 12 )kii ) > 0}
i=1
{ak11bak22b. . . :
∞
(1 − ( 12 )ki ) = 0}
98 / 148
PBA with standard and alternative semantics
pba-20
PBA with thresholds (a+b)∗aω
{ak11bak2 b. . . :
PBA, PSA, PRA and 0/1-PRA ω -regular languages uniform PBA PBA almost sure semantics DBA ∞
(1 − ( 12 )kii ) > 0}
i=1
{ak11bak22b. . . :
∞
(1 − ( 12 )ki ) = 0}
99 / 148
0/1-PRA
pba-20b
0/1-PRA: probabilistic Rabin automaton P s.t. ∀x ∈ Σω : PrP (x) ∈ {0, 1}
100 / 148
0/1-PRA
pba-20b
0/1-PRA: probabilistic Rabin automaton P s.t. ∀x ∈ Σω : PrP (x) ∈ {0, 1} Theorem: For each (standard) PBA P there exists a P ) = L (P PR ). 0/1-PRA PR such that L (P
101 / 148
0/1-PRA
pba-20b
0/1-PRA: probabilistic Rabin automaton P s.t. ∀x ∈ Σω : PrP (x) ∈ {0, 1} Theorem: For each (standard) PBA P there exists a P ) = L (P PR ). 0/1-PRA PR such that L (P Corollary: The almost-sure semantics for PRA is as powerful as the standard semantics. The same holds for PSA, but not for PBA.
102 / 148
0/1-PRA
pba-20b
0/1-PRA: probabilistic Rabin automaton P s.t. ∀x ∈ Σω : PrP (x) ∈ {0, 1} Theorem: For each (standard) PBA P there exists a P ) = L (P PR ). 0/1-PRA PR such that L (P idea: 0/1-PRA PR • generates up to n = |Q| sample runs of P (as representatives for all runs in P ) • and checks whether at least one of them is accepting
103 / 148
From PBA P to 0/1-PRA P R pairwise distinct states in P
pba-20b
powerset construction
state in P R : q1, q2 , . . . , qi , R
104 / 148
From PBA P to 0/1-PRA P R pairwise distinct states in P state in P R : q1, q2 , . . . , qi , R a a ... a q1 , q2, . . . , qi ,
δ
pba-20b
powerset construction
a δ (R, a)
= transition function in P 105 / 148
From PBA P to 0/1-PRA P R pairwise distinct states in P state in P R : q1, q2 , . . . , qi , R a a ... a
pba-20b
powerset construction
a
. . . qk , δ (R, a) q1 , q2, . . . , qi , qi+1
where {qi+1 , . . . , qk } = F ∩ δ (R, a)\{q1 , . . . , qi }
F = set of accept states in P δ
= transition function in P 106 / 148
From PBA P to 0/1-PRA P R pairwise distinct states in P state in P R : q1, q2 , . . . , qi , R a a ... a
pba-20b
powerset construction
a
. . . qk , δ (R, a) q1 , q2, . . . , qi , qi+1
, . . . , qk , δ (R, a) next state q1 , q3 , . . . qi ,qi+1
shift
107 / 148
From PBA P to 0/1-PRA P R pairwise distinct states in P state in P R : q1, q2 , . . . , qi , R a a ... a
pba-20b
powerset construction
a
. . . qk , δ (R, a) q1 , q2, . . . , qi , qi+1
, . . . , qk , δ (R, a) next state q1 , q3 , . . . qi ,qi+1
shift
acceptance condition: ♦ “no shift in component j” ∧ ♦ F) (♦ j
108 / 148
Complementation of PBA
pba-21
Theorem: The class of PBA-recognizable languages is closed under complementation.
109 / 148
Complementation of PBA
pba-21
Theorem: The class of PBA-recognizable languages is closed under complementation.
PBA P
110 / 148
Complementation of PBA
pba-21
Theorem: The class of PBA-recognizable languages is closed under complementation.
PBA −→ 0/1-PRA P
PR
111 / 148
Complementation of PBA
pba-21
Theorem: The class of PBA-recognizable languages is closed under complementation. compl.
PBA −→ 0/1-PRA −−−→ 0/1-PSA P
PR
112 / 148
Complementation of PBA
pba-21
Theorem: The class of PBA-recognizable languages is closed under complementation. compl.
PBA −→ 0/1-PRA −−−→ 0/1-PSA P
PR
=
PS
113 / 148
Complementation of PBA
pba-21
Theorem: The class of PBA-recognizable languages is closed under complementation. compl.
PBA −→ 0/1-PRA −−−→ 0/1-PSA −→ PBA P
PR
=
PS
P
114 / 148
Complementation of PBA
pba-21
Theorem: The class of PBA-recognizable languages is closed under complementation. exp
compl.
poly
PBA −−→ 0/1-PRA −−−→ 0/1-PSA −−→ PBA P
PR
=
PS
P
115 / 148
Operators on PBA
pba-21
Theorem: The class of PBA-recognizable languages is closed under complementation, union, intersection. complementation: exp
compl.
poly
PBA −−→ 0/1-PRA −−−→ 0/1-PSA −−→ PBA union and intersection: • union: random choice between two PBA • intersection: via generalized PBA (as for NBA)
116 / 148
Decision problems for PBA
pba-22
The emptiness problem for PBA given: PBA P P ) = ∅ hold? question: does L (P
117 / 148
Decision problems for PBA
pba-22
The emptiness problem for PBA given: PBA P P ) = ∅ hold? question: does L (P is undecidable.
118 / 148
Decision problems for PBA
pba-22
The emptiness problem for PBA given: PBA P P ) = ∅ hold? question: does L (P is undecidable. proof by a reduction from the emptiness problem for probabilistic finite automata (PFA) [Paz’71, Madani/Hanks/Condon’03]
119 / 148
Decision problems for PBA
pba-22
The emptiness problem for PBA is undecidable. Hence, the following problems are undecidable too:
120 / 148
Decision problems for PBA
pba-22
The emptiness problem for PBA is undecidable. Hence, the following problems are undecidable too: • universality: given a PBA P , P ) = Σω hold? does L (P
121 / 148
Decision problems for PBA
pba-22
The emptiness problem for PBA is undecidable. Hence, the following problems are undecidable too: • universality: given a PBA P , P ) = Σω hold? does L (P • equivalence: given two PBA P 1 , P 2 , P 1 ) = L (P P 2 ) hold? does L (P
122 / 148
Decision problems for PBA
pba-22
The emptiness problem for PBA is undecidable. Hence, the following problems are undecidable too: P ) = Σω hold? • universality: does L (P P 1 ) = L (P P 2 ) hold? • equivalence: does L (P • model checking finite transition systems against PBA-specifications
123 / 148
Decision problems for PBA
pba-22
The emptiness problem for PBA is undecidable. Hence, the following problems are undecidable too: P ) = Σω hold? • universality: does L (P P 1 ) = L (P P 2 ) hold? • equivalence: does L (P • model checking finite transition systems against PBA-specifications given a finite TS T and PBA P , does there exist P )? a path π in T s.t. trace(π) ∈ L (P
124 / 148
Decision problems for PBA
pba-22
The emptiness problem for PBA is undecidable. Hence, the following problems are undecidable too: P ) = Σω hold? • universality: does L (P P 1 ) = L (P P 2 ) hold? • equivalence: does L (P • model checking finite transition systems against PBA-specifications given a finite TS T and PBA P , does there exist P )? a path π in T s.t. trace(π) ∈ L (P given a finite TS T and PBA P , does P ) hold for all paths π in T ? trace(π) ∈ L (P 125 / 148
Decision problems for PBA
pba-22
The following problems are undecidable: • emptiness for PBA • universality for PBA • equivalence for PBA • model checking finite TS against PBA-specifications • verification of observation-based stochastic games formalized by partially-observable MDPs (POMDPs) against ω -regular specifications:
126 / 148
Decision problems for PBA
pba-22
The following problems are undecidable: • emptiness for PBA • universality for PBA • equivalence for PBA • model checking finite TS against PBA-specifications • verification of observation-based stochastic games formalized by partially-observable MDPs (POMDPs) against ω -regular specifications: ♦ F) > 00? * does there exist a strategy S s.t. Pr S ( ♦ F) = 11? * does there exist a strategy S s.t. Pr S(♦ 127 / 148
Decision problems for POMDPs
pba-22a
Theorem: The following instances of the verification problem for finite POMDPs against ω -regular specifications is undecidable: ♦ F) > 00? * does there exist a strategy S s.t. Pr S( ♦ F) = 11? * does there exist a strategy S s.t. Pr S(♦
128 / 148
Decision problems for POMDPs
pba-22a
Theorem: The following instances of the verification problem for finite POMDPs against ω -regular specifications is undecidable: ♦ F) > 00? * does there exist a strategy S s.t. Pr S( ♦ F) = 11? * does there exist a strategy S s.t. Pr S(♦ but the following problems are decidable: F) > 00? * does there exist a strategy S s.t. Pr S( [de Alfaro’99]
129 / 148
Decision problems for POMDPs
pba-22a
Theorem: The following instances of the verification problem for finite POMDPs against ω -regular specifications is undecidable: ♦ F) > 00? * does there exist a strategy S s.t. Pr S( ♦ F) = 11? * does there exist a strategy S s.t. Pr S(♦ but the following problems are decidable: F) > 00? * does there exist a strategy S s.t. Pr S( [de Alfaro’99] ♦F) = 11? * does there exist a strategy S s.t. Pr S(♦ 130 / 148
Decision problems for POMDPs
pba-22a
Theorem: The following instances of the verification problem for finite POMDPs against ω -regular specifications is undecidable: ♦ F) > 00? * does there exist a strategy S s.t. Pr S( ♦ F) = 11? * does there exist a strategy S s.t. Pr S(♦ but the following problems are decidable: F) > 00? * does there exist a strategy S s.t. Pr S( ♦F) = 11? * does there exist a strategy S s.t. Pr S(♦ ♦ F) = 11? * does there exist a strategy S s.t. Pr S( ♦ F) > 00? * does there exist a strategy S s.t. Pr S(♦ 131 / 148
Decision problems for POMDPs
pba-22a
Theorem: The following instances of the verification problem for finite POMDPs are decidable: .. . ♦ F) = 11? does there exist a strategy S s.t. Pr S ( .. .
132 / 148
Decision problems for POMDPs
pba-22a
Theorem: The following instances of the verification problem for finite POMDPs are decidable: .. . ♦ F) = 11? does there exist a strategy S s.t. Pr S ( .. . Corollary: The emptiness problem for PBA with the almost-sure semantics is decidable.
133 / 148
... undecidability results for general PBA, but ...
pba-23
• emptiness problem for PBA with the almost-sure semantics is in EXPTIME
134 / 148
... undecidability results for general PBA, but ...
pba-23
• emptiness problem for PBA with the almost-sure semantics is in EXPTIME • emptiness problem for uniform PBA is in EXPTIME
135 / 148
... undecidability results for general PBA, but ...
pba-23
• emptiness problem for PBA with the almost-sure semantics is in EXPTIME • emptiness problem for uniform PBA is in EXPTIME exp
via uPBA −−→ NSA
136 / 148
... undecidability results for general PBA, but ...
pba-23
• emptiness problem for PBA with the almost-sure semantics is in EXPTIME • emptiness problem for uniform PBA is in EXPTIME • model checking Markov chains against uPBA-spec given: finite Markov chain M , uniform PBA P L(P P )) > 0 hold? question: does PrM (L
137 / 148
... undecidability results for general PBA, but ...
pba-23
• emptiness problem for PBA with the almost-sure semantics is in EXPTIME • emptiness problem for uniform PBA is in EXPTIME • model checking Markov chains against uPBA-spec given: finite Markov chain M , uniform PBA P L(P P )) > 0 hold? question: does PrM (L L(P P )) denotes the probability measure where PrM (L π ) ∈ L (P P) of the set of paths π in M s.t. trace(π 138 / 148
... undecidability results for general PBA, but ...
pba-23
• emptiness problem for PBA with the almost-sure semantics is in EXPTIME • emptiness problem for uniform PBA is in EXPTIME • model checking Markov chains against uPBA-spec given: finite Markov chain M , uniform PBA P L(P P )) > 0 hold? question: does PrM (L is in PTIME as L(P P )) > 0 iff PrM×P ( ♦ F) > 0 PrM (L
139 / 148
... undecidability results for general PBA, but ...
pba-23
• emptiness problem for PBA with the almost-sure semantics is in EXPTIME • emptiness problem for uniform PBA is in EXPTIME • model checking Markov chains against uPBA-spec given: finite Markov chain M , uniform PBA P L(P P )) > 0 hold? question: does PrM (L is in PTIME as L(P P )) > 0 iff PrM×P ( ♦ F) > 0 PrM (L ↑ analysis of the SCCs in the product Markov chain 140 / 148
Conclusion
pba-conc
• expressiveness: PBA are more expressive than NBA and almost-sure PBA,
141 / 148
Conclusion
pba-conc
• expressiveness: PBA are more expressive than NBA and almost-sure PBA, while uniform PBA have equal power than NBA
142 / 148
Conclusion
pba-conc
• expressiveness: PBA are more expressive than NBA and almost-sure PBA, while uniform PBA have equal power than NBA • efficiency: PBA can be exp smaller than NSA
143 / 148
Conclusion
pba-conc
• expressiveness: PBA are more expressive than NBA and almost-sure PBA, while uniform PBA have equal power than NBA • efficiency: PBA can be exp smaller than NSA • polynomial transformation PSA PBA
144 / 148
Conclusion
pba-conc
• expressiveness: PBA are more expressive than NBA and almost-sure PBA, while uniform PBA have equal power than NBA • efficiency: PBA can be exp smaller than NSA • polynomial transformation PSA PBA • undecidability results for PBA and POMDPs
145 / 148
Conclusion
pba-conc
• expressiveness: PBA are more expressive than NBA and almost-sure PBA, while uniform PBA have equal power than NBA • efficiency: PBA can be exp smaller than NSA • polynomial transformation PSA PBA • undecidability results for PBA and POMDPs • decidability results for almost-sure PBA and POMDPs
146 / 148
Conclusion
pba-conc
• expressiveness: PBA are more expressive than NBA and almost-sure PBA, while uniform PBA have equal power than NBA • efficiency: PBA can be exp smaller than NSA • polynomial transformation PSA PBA • undecidability results for PBA and POMDPs • decidability results for almost-sure PBA and POMDPs • application: run-time verification (probabilistic monitoring) [Sistla et al] 147 / 148
Conclusion • • • • •
pba-conc
expressiveness ... efficiency .... polynomial transformation PSA PBA (un)decidability results for PBA and POMDPs application: run-time verification [Sistla et al]
many open problems: • transformations LTL PBA or MSO PBA • alternative semantics for PBA • variants: NPBA, QBA, ... 148 / 148