Download as pdf or txt
Download as pdf or txt
You are on page 1of 48

STIT fragment

1
Introduction

Strategic fragment of Losstit

Normal simulation of monotonic modal logic and Kracht-Wolter formulas

Bridges between Lcstit and Losstit

2
Introduction
Motivation

STIT (sees to it that) semantics is one of the most prominent tools in modal logic of
agency, widely used among both philosophers and responsible AI scholars. STIT logic
surveys the properties of agents seeing to it that some state of affairs holds without
specifying concrete actions by which that state of affairs is guaranteed. In comparison
with other multi-agent modal logics, the main advantage of STIT theories is expressive
power. STIT logic allows to study not only statements about agents’ abilities to perform
certain actions (as it is in variations of Coalition Logic or Propositional Dynamic Logic),
but about what choices they make and what they de-facto achieve as well.

3
Motivation

STIT (sees to it that) semantics is one of the most prominent tools in modal logic of
agency, widely used among both philosophers and responsible AI scholars. STIT logic
surveys the properties of agents seeing to it that some state of affairs holds without
specifying concrete actions by which that state of affairs is guaranteed. In comparison
with other multi-agent modal logics, the main advantage of STIT theories is expressive
power. STIT logic allows to study not only statements about agents’ abilities to perform
certain actions (as it is in variations of Coalition Logic or Propositional Dynamic Logic),
but about what choices they make and what they de-facto achieve as well.
Nevertheless, in some occasions such expressivity may be redundant.

3
Classic STIT logic

The language of ”classic” stit logic Lcstit is defined as follows. For a countable set of
propositional variables Var = {p1 , p2 , . . .} and a finite set of agents Ags = {i1 , i2 , . . . , in }::

ϕ := p | ¬ϕ | ϕ ∨ ϕ | 2ϕ | [cstit]i ϕ

where p ∈ Var, i ∈ Ags. [cstit]i ϕ – «agent i sees to it that ϕ ». 2ϕ – «historically necessary,


that ϕ ». Standardly, hcstitii ϕ ≡ ¬[cstit]i ¬ϕ ; 3ϕ ≡ ¬2¬ϕ .

4
Kripke semantics for Lcstit

We will present a simplified Kripke semantics: (balb).

5
Kripke semantics for Lcstit

We will present a simplified Kripke semantics: (balb).

M = hW, R, {Ri }i∈Ags , Vi

• W 6= 0/ – non-empty set of possible worlds

5
Kripke semantics for Lcstit

We will present a simplified Kripke semantics: (balb).

M = hW, R, {Ri }i∈Ags , Vi

• W 6= 0/ – non-empty set of possible worlds


• R ⊆ W × W – equivalency relation of historical accessibility, defined of W

5
Kripke semantics for Lcstit

We will present a simplified Kripke semantics: (balb).

M = hW, R, {Ri }i∈Ags , Vi

• W 6= 0/ – non-empty set of possible worlds


• R ⊆ W × W – equivalency relation of historical accessibility, defined of W
• Ri ⊆ W × W – equivalency relation of agent’s i efficiency, defined on W. If w1 Ri w2 , then w1 , w2 –
potential outcomes of some agent’s i action

5
Kripke semantics for Lcstit

We will present a simplified Kripke semantics: (balb).

M = hW, R, {Ri }i∈Ags , Vi

• W 6= 0/ – non-empty set of possible worlds


• R ⊆ W × W – equivalency relation of historical accessibility, defined of W
• Ri ⊆ W × W – equivalency relation of agent’s i efficiency, defined on W. If w1 Ri w2 , then w1 , w2 –
potential outcomes of some agent’s i action
• Ri ⊆ R
• Independence of agents:
∀w, v ∈ W ∀a, b, c ∈ Ags((w, v) ∈ Ra ◦ Rb → ∃u ∈ W ∀i ∈ Ags \ {c}(w, u) ∈ Rc ∧ (u, v ∈ Ri )).
Informally: agents are unable to simultaneously execute mutually exclusive actions.
• V : Var → 2W – standard evaluation function

5
Independence
a b of agents

b
a
v w
u a z

b b
a
6
v w
Independence
u of agents

b
a
v w
u a z

b b
a
v w

7
Kripke semantics for Lcstit

M , w |= p ⇔ w ∈ V(p)
M , w |= ¬ϕ ⇔ M , w 6|= ϕ
M , w |= ϕ ∨ ψ ⇔ M , w |= ϕ or M , w |= ψ
M , w |= 2ϕ ⇔ ∀w0 ∈ W(wRw0 → M , w0 |= ϕ )
M , w |= [cstit]i ϕ ⇔ ∀w0 ∈ W(wRi w0 → M , w0 |= ϕ )

8
Kripke semantics for Lcstit

M , w |= p ⇔ w ∈ V(p)
M , w |= ¬ϕ ⇔ M , w 6|= ϕ
M , w |= ϕ ∨ ψ ⇔ M , w |= ϕ or M , w |= ψ
M , w |= 2ϕ ⇔ ∀w0 ∈ W(wRw0 → M , w0 |= ϕ )
M , w |= [cstit]i ϕ ⇔ ∀w0 ∈ W(wRi w0 → M , w0 |= ϕ )

Let G ⊆ Ags. In a richer language, we could speak of group actions:



M , w |= [cstit]G ϕ ⇔ ∀w0 ∈ W((w, w0 ) ∈ Ri → M , w0 |= ϕ )
i∈G

8
Lcstit logic

If we do not have group actions, Lcstit is sound and strongly complete w.r.t. a class of
Kripke frames defined earlier (balb):

(PL) Tautologies of classical propositional logic


(S52) S5 for 2
(S5 [cstit]) S5 for [cstit]i
(Nec) 2ϕ → [cstit]i ϕ
(Ind) 3[cstit]1 ϕ1 ∧ . . . ∧ 3[cstit]n ϕn → 3([cstit]1 ϕ1 ∧ . . . ∧ [cstit]n ϕn )

9
Abilities in Lcstit and why do we need this fragment

We could speak of not only what was de-facto achieve – [cstit]i ϕ , but about strategic
abilities as well: 3[cstit]i ¬ϕ ∧ [cstit]i ϕ . Same expressions could be found in other
multi-agent modal logics: 3[cstit]i ϕ may be seen as «synonym» h[{i}]iϕ in Coalition logic.
In some occasions in may be useful to complete the same analysis inside STIT theory.

10
Abilities in Lcstit and why do we need this fragment

We could speak of not only what was de-facto achieve – [cstit]i ϕ , but about strategic
abilities as well: 3[cstit]i ¬ϕ ∧ [cstit]i ϕ . Same expressions could be found in other
multi-agent modal logics: 3[cstit]i ϕ may be seen as «synonym» h[{i}]iϕ in Coalition logic.
In some occasions in may be useful to complete the same analysis inside STIT theory.

Nevertheless, the big problem with STIT-logic is its computational properties: Lcstit with
group operators [cstit]G ϕ neither finitely axiomatisable, nor decidable in case of |Ags| > 3
(Herzig and Schwarzentruber 2008). It is also known that SAT problem for Lcstit without
group operators is NEXPTIME-complete in case of |Ags| > 2 (balb). That may be another
motivation for research of STIT fragments.

10
Strategic fragment of Losstit
The language of Losstit

For a countable set of propositional letters Var = {p1 , p2 , . . .} and finite set of agents
Ags = {i1 , i2 , . . . , in }, BNF for Losstit syntax:

ϕ := p | ¬ϕ | ϕ ∨ ϕ | 2ϕ | [i]ϕ | [∃i ]ϕ

where p ∈ Var, i ∈ Ags.


[i]ϕ – «agent i has a strategy to see to it that ϕ » (analogous to 3[cstit]i ϕ );
[∃i ]ϕ – «agent i has no strategy to prevent ϕ » (analogous to 2[cstit]ϕ );

11
Neighbourhood semantics for Losstit

M = hW, {Choicei }i∈Ags , Vi

• W 6= 0/ – non-empty set of possible worlds, all of them are historically accessible

12
Neighbourhood semantics for Losstit

M = hW, {Choicei }i∈Ags , Vi

• W 6= 0/ – non-empty set of possible worlds, all of them are historically accessible


W
• Choicei : W → 22 – neighbourhood function, defined for every agent
• V : Var → 2W – standard evaluation function

12
Constraints on Choice

1. ∀w ∈ W : 0/ 6∈ Choicei (w)

13
Constraints on Choice

1. ∀w ∈ W : 0/ 6∈ Choicei (w)
2. ∀w ∈ W : W ∈ Choicei (w)

13
Constraints on Choice

1. ∀w ∈ W : 0/ 6∈ Choicei (w)
2. ∀w ∈ W : W ∈ Choicei (w)
3. ∀w ∈ W ∀X, Y ⊆ W : (X ∈ Choicei (w) ∧ X ⊆ Y) → Y ∈ Choicei (w)

13
Constraints on Choice

1. ∀w ∈ W : 0/ 6∈ Choicei (w)
2. ∀w ∈ W : W ∈ Choicei (w)
3. ∀w ∈ W ∀X, Y ⊆ W : (X ∈ Choicei (w) ∧ X ⊆ Y) → Y ∈ Choicei (w)

4. Choicei ⇂ (w) = W (Choicei ⇂ – is just Choice without supersets)

13
Constraints on Choice

1. ∀w ∈ W : 0/ 6∈ Choicei (w)
2. ∀w ∈ W : W ∈ Choicei (w)
3. ∀w ∈ W ∀X, Y ⊆ W : (X ∈ Choicei (w) ∧ X ⊆ Y) → Y ∈ Choicei (w)

4. Choicei ⇂ (w) = W (Choicei ⇂ – is just Choice without supersets)
5. ∀w ∈ W ∀X, Y ⊆ W : (X ∈ Choicea (w) ∧ Y ∈ Choiceb (w)) → X ∩ Y 6= 0/

13
Constraints on Choice

1. ∀w ∈ W : 0/ 6∈ Choicei (w)
2. ∀w ∈ W : W ∈ Choicei (w)
3. ∀w ∈ W ∀X, Y ⊆ W : (X ∈ Choicei (w) ∧ X ⊆ Y) → Y ∈ Choicei (w)

4. Choicei ⇂ (w) = W (Choicei ⇂ – is just Choice without supersets)
5. ∀w ∈ W ∀X, Y ⊆ W : (X ∈ Choicea (w) ∧ Y ∈ Choiceb (w)) → X ∩ Y 6= 0/
6. ∀w ∈ W X ⊆ W : X ∈ Choicei (w) → ∀w0 ∈ W : X ∈ Choicei (w0 )

13
Neighbourhood semantics Losstit

M , w |= p ⇔ w ∈ V(p)
M , w |= ¬ϕ ⇔ M , w 6|= ϕ
M , w |= ϕ ∨ ψ ⇔ M , w |= ϕ or M , w |= ψ
M , w |= 2ϕ ⇔ ∀w0 ∈ W(M , w0 |= ϕ )
M , w |= [i]ϕ ⇔ Jϕ K ∈ Choicei (w)
M , w |= [∃i ]ϕ ⇔ ∀X ∈ Choicei ⇂ (w)∀w0 ∈ X(M , w0 |= ϕ )

14
Neighbourhood semantics Losstit

M , w |= p ⇔ w ∈ V(p)
M , w |= ¬ϕ ⇔ M , w 6|= ϕ
M , w |= ϕ ∨ ψ ⇔ M , w |= ϕ or M , w |= ψ
M , w |= 2ϕ ⇔ ∀w0 ∈ W(M , w0 |= ϕ )
M , w |= [i]ϕ ⇔ Jϕ K ∈ Choicei (w)
M , w |= [∃i ]ϕ ⇔ ∀X ∈ Choicei ⇂ (w)∀w0 ∈ X(M , w0 |= ϕ )
Non-monotonic definition for [i]ϕ :

M , w |= [i]ϕ ⇔ ∃X ∈ Choicei ⇂ (w) ∀w0 ∈ X(M , w |= ϕ )

14
Logic Losstit

(PL) All classical tautologies


(S52) S5 for 2
(S5[i]) S5 for [∃i ]
(Incl) 2ϕ → [i]ϕ
(M) [i](ϕ ∧ ψ ) → ([i]ϕ ∧ [i]ψ )
(N) [i]>
(D) ¬[i]⊥
(Pos) 2ϕ ≡ [∃i ]ϕ
(Nec-A) [i]ϕ → 2[i]ϕ
(Ind) [1]ϕ1 ∧ [2]ϕ2 ∧ . . . ∧ [n]ϕn → 3(ϕ1 ∧ ϕ2 ∧ . . . ∧ ϕn )
(RE) From ϕ ≡ ψ , infer [i]ϕ ≡ [i]ψ
(RE) From ϕ ≡ ψ , infer [∃i ]ϕ ≡ [∃i ]ψ
(MP) From ϕ , ϕ → ψ , infer ψ

15
Logic Losstit

(PL) All classical tautologies


(S52) S5 for 2
(S5[i]) S5 for [∃i ]
(Incl) 2ϕ → [i]ϕ
(M) [i](ϕ ∧ ψ ) → ([i]ϕ ∧ [i]ψ )
(N) [i]>
(D) ¬[i]⊥
(Pos) 2ϕ ≡ [∃i ]ϕ
(Nec-A) [i]ϕ → 2[i]ϕ
(Ind) [1]ϕ1 ∧ [2]ϕ2 ∧ . . . ∧ [n]ϕn → 3(ϕ1 ∧ ϕ2 ∧ . . . ∧ ϕn )
(RE) From ϕ ≡ ψ , infer [i]ϕ ≡ [i]ψ
(RE) From ϕ ≡ ψ , infer [∃i ]ϕ ≡ [∃i ]ψ
(MP) From ϕ , ϕ → ψ , infer ψ

15
Soundness and strong completeness

The proof of soundness and strong completeness is obtained either by canonical model
construction or automatically, since all axioms, except (M) and (D), are Kracht-Wolter
formulas

16
Normal simulation of monotonic
modal logic and Kracht-Wolter
formulas
Kracht-Wolter formulas

Assume a modal language L1 , which is semantically defined on a monotonic


neighbourhood frames. Then Kracht-Wolter formulas are implications of the form τ → χ ,
such that τ is formulated as

∧ ∧
2 ϕi ∧ ψj
i≤n j≤m

and χ is constructed from meta-variables and ∧, ∨, 2, 3 only.

17
Kracht-Wolter formulas

Assume a modal language L1 , which is semantically defined on a monotonic


neighbourhood frames. Then Kracht-Wolter formulas are implications of the form τ → χ ,
such that τ is formulated as

∧ ∧
2 ϕi ∧ ψj
i≤n j≤m

and χ is constructed from meta-variables and ∧, ∨, 2, 3 only. If ϕ is a Kracht-Wolter


formula (kracht), then it is possible to efficiently compute a first-order constraint on

neighbourhood frames α (ϕ ), such that a logic M ϕ is canonical w.r.t. a class of
neighbourhood frames satisfying α (ϕ ).

17
Strong completeness of Losstit

(Almost) all axioms of Losstit are Kracht-Wolter formulas, corresponding to first-order


conditions on Choice functions:

(M) Monotonicity of Ni
(C) ∀w ∈ W∀X, Y ⊆ W((X ∈ Ni (w) ∧ Y ∈ Ni (w)) → X ∩ Y ∈ Ni (w))
(N) W ∈ Ni (w)
(D) 0/ 6∈ Ni (w)
(T) ∀w ∈ W(w ∈ Ni (w))
(4’) ∀w ∈ W X ⊆ W(X ∈ Ni (w) → N−1 i
(X) ∈ Ni (w))
(B) ∀w ∈ W X ⊆ W(w ∈ X → W \ (N−1 i
(W \ X) ∈ Ni (w)))

(Pos) Ni ⇂ (w) = W
(Nec-A) Ni (w) = Ni (w0 )
(Ind) ∀X ∈ Na (w), Y ∈ Nb (w)(X ∩ Y 6= 0)/

18
Completeness Losstit

The only axiom which is not a Kracht-Wolter formula is (D): ¬[i]⊥. Nevertheless, it is
known to be canonical, and we still can construct a canonical model, satisfying (D)
together with other axioms, such that our logic will be valid on the model
(khaitovich2022neighbourhood).

19
Bridges between Lcstit and Losstit
Semantics for 3[cstit]i ϕ vs. [i]ϕ

M , w |= 3[cstit]i ϕ ⇔ ∃w0 ∈ R(w)∀w00 ∈ Ri (w0 )(M , w00 |= ϕ )


This definition completely identical with non-monotonic definition for [i]ϕ .

20
What is the difference?

So we have defined the language Losstit , some class of neighbourhood frames and logic
Losstit , after what we have proven completeness of the latter w.r.t. that class of frames.
But why it is a fragment of Lcstit ?

21
What is the difference?

So we have defined the language Losstit , some class of neighbourhood frames and logic
Losstit , after what we have proven completeness of the latter w.r.t. that class of frames.
But why it is a fragment of Lcstit ?

It is worth noticing that there are 2 main differences between Lcstit Kripke frames and
neighbourhood frames of Losstit :

1. Efficiency relation Ri (w) partitions a set of historically accessible worlds R(w), while
Choicei (w) do not partition W.
2. In Kripke frames, historical accessibility relation R is not total

21
Neighbourhoods as partitions

Assume a class of neighbourhood frames P, such that for all frames in this class
F = hW, Ni and for all w ∈ W, N ⇂ (w) is a partition of W, i.e.

1. N(w) = W
2. 0/ 6∈ N(w)
3. ∀X, Y ∈ N(w)(X ∩ Y = 0)
/

22
Modal undefinability of P

Let F1 = hW1 , N1 i and F2 = hW2 , N2 i be two neighbourhood frames, such that:

1. W1 = {w1 , w2 , w3 , w4 }, N1 ⇂ (w1 ) = {{w1 , w2 }, {w3 , w4 }}, while for other worlds w2 , . . . , w4 ,


N1 ⇂ (w) = {W1 }. Obviously, F1 ∈ P
2. W2 = {w1 , w2 , w3 }, N2 ⇂ (w1 ) = {{w1 , w2 }, {w2 , w3 }} while for other worlds w2 , . . . , w4 ,
N2 ⇂ (w) = {W2 }.Obviously, F2 6∈ P
3. Let f : W1 → W2 be a surjective, function, such that
f : w1 7→ w1 ; f : w2 7→ w2 ; f : w3 7→ w3 ; f : w4 7→ w2 ;
4. f is a surjective p-morphism from W1 to W2 hansen, pac. For all w ∈ W1 :
4.1 If X ∈ N1 (w), then f[X] ∈ N2 (f(w))
4.2 If Y ∈ N2 (f(w)), then there is a X ⊆ W1 , such that f[X] = Y and X ∈ N1 (w)

23
Disjoint unions

Notice that in case of Losstit frames W is a set of historically accessible worlds and R is a
universal relation. While in Lcstit it is not the case: R partitions W on equivalence classes,
where every class may be seen as a set of historically accessible worlds. Moreover,
possible worlds from different equivalence classes are not connected with each other by
any other relation.

24
Disjoint unions

Notice that in case of Losstit frames W is a set of historically accessible worlds and R is a
universal relation. While in Lcstit it is not the case: R partitions W on equivalence classes,
where every class may be seen as a set of historically accessible worlds. Moreover,
possible worlds from different equivalence classes are not connected with each other by
any other relation.

That is why Lcstit frames may be seen as a disjoint unionы of sets of Losstit frames. Since
disjoint unions preserve modal similarity, Losstit logic will still do its job,

24
Further research

• Try to do the same for epistemic extensions: it is not immediately clear if an


epistemic possibility operator [Ki ]ϕ (agent i is able to knowingly see to it that ϕ )? Is it
appropriate to define such new operator or we can obtain the same results by
combination of [i] and Ki ?
• Explore the same fragment for group STIT
• Survey computational issues

25
References i

refs

26

You might also like