Lecture 4:

Teachers: Dietmar Berwanger

Construction of a winning strategy in a finite parity game

• pick position $q$ of minium priority (say even)
• Induction Hypothesis: $V \backslash Attr^0(\lbrace q \rbrace)$ splits into winning regions $(W_0’, W_1’)$
• $W_1’ ⊆ W_1$
• How about $W_0’ ∪ Attr^0(\lbrace q\rbrace)$?

• $q ∈ V_0$ and $qE ⊆ W_1’$ OR $q ∈ V_1$ and $qE ∩ W_1’ ≠ ∅$:

By IH: split $V \backslash (Attr^1(\lbrace q \rbrace) ∪ W_1’)$ into $(W_0’’, W_1’’)$.

In this case: $W_0’ = W_0, \; W_1 = Attr^1(\lbrace q \rbrace) ∪ W_1’ ∪ W_1’’$: indeed, it suffices to show:

• $W_0’ ⊆ W_0$
• $Attr^1(\lbrace q \rbrace) ∪ W_1’ ∪ W_1’’ ⊆ W_1$
• $q ∈ V_0$ and $qE ∩ (W_0’ ∪ Attr^0(\lbrace q\rbrace)) ≠ ∅$ OR $q ∈ V_1$ and $qE ⊆ W_0’ ∪ Attr^0(\lbrace q\rbrace)$: we’re done, P0 wins

Thm (Parity-Decidability): “Given $G, v$ decide whether P0 has winning strategy from $v$” is in NP.

Given a memoryless strategy $σ:V_0 ⟶ V$ we can decide in PTIME whether it is winning.

 $G_{ σ} ≝ (V, E_σ, V_0, V_1)$ with $E_σ$ st
• for $v ∈ V_1$, $vE_σ = vE$
• for $v ∈ V_0$, $vE_σ = σ(v)$
 $σ$ winning ⟺ in $G_{ σ}$, all reachable from $v_0$ cycles have an even minimum.

And this property can be decided in time $O\Big(\frac {\vert Ω(V) \vert (\vert V \vert + \vert E \vert)} 2\Big)$ for a graph $G ≝ (V, E, Ω)$:

• Idea: for odd priorities: $1, 3, 5, …, m$:

consider

• $G_k ≝ G_{| \lbrace v ∈ V \; \mid \; Ω(v) ≥ k\rbrace}$
• decompose $G_k$ into strong connected components (SCC) (Tarjan algorithm)
• check existence of a SCC “reachable from $v_0$ in $G$” that contains $w$ with $Ω(w) = k$: if yes, reject

else: repeat with another value of $k$

$ParityDcc ∈ NP ∩ coNP$

Application

Synthesis problem:

• input: $\overline{x} ∈ Σ^ω$
• output: $\overline{y} ∈ \Pi^ω$

Specification: Relation

$R ⊆ (Σ × \Pi)^ω$

Question: contruct a function $f: Σ^\ast → \Pi$ st

$(x_1 x_2 ⋯, f(ε) f(x_1) f(x_1 x_2) ⋯) ∈ R$

Indeed: Here $R = ℒ(A)$ (regular relation, when representable by a 2-tape automaton) for

$𝒜$ a 2-tape automaton:
$(Q, Σ × \Pi, q_0, δ: Q × (Σ × \Pi) → Q, Acc)$

Acceptability condition: Reach, Büchi (better for non-determinism), Parity

Variants: decide whether a (possibly computable/finite-state/automaton) solution $f$ exists

Regular function:

$(Q, \underbrace{Σ}_{\text{input}}, \underbrace{\Pi}_{\text{output}}, q_0, δ: Q × Σ → Q × \Pi)$ or Mealy machine: $(Q, Σ, \Pi, q_0, δ: Q × Σ → Q, \underbrace{β: Q × Σ → \Pi}_{\text{generation of output}})$ or Moore machine: $(Q, Σ, \Pi, q_0, δ: Q × Σ → Q, ζ: Q → \Pi)$

Extending transition output for

$δ: Q × Σ^\ast → Q\\ δ(q,ε) = q\\ δ(q, w a) = δ(δ(q,w),a)$ where $q ∈ Q, a ∈ Σ, w ∈ Σ^\ast$

Similarly for the output function:

$β: Q × Σ^\ast → \Pi\\ β(q,ε) = q\\ β(q, w a) = β(δ(q,w),a)$

Back to the problem: given $Σ, \Pi, 𝒜$ with $ℒ(A) = Spec ⊆ (Σ × \Pi)^ω$.

We want $ℳ ≝ (M, Σ × \Pi, m_0, ν: M × E → M, β: M× E → \Pi)$ st

$x_1, x_2 …, β(q_0, x_1), β(q_0, x_1 x_2) … β(q_0, x_1 ⋯ x_t) ∈ ℒ(𝒜)$
  digraph {
rankdir=LR;
invis1[shape=none, label=""];
invis2[shape=rectangle, label=""];
invis3[shape=none, label=""];
invis1 -> invis2[label="Σ"];
invis2 -> invis3[label="Π"];
}

1. Represent the system as a parity game: $G ≝ (V, V_0, E, Ω), v_0$

with

$V = Q × (Σ ∪ \Pi) ∪ \lbrace q_0\rbrace\\ V_0 = Q × Σ \qquad V_1 = Q × \Pi ∪ \lbrace q_0\rbrace$

Moves $E$:

$E ≝ \lbrace (⟨q, a⟩, \big\langle δ(q, ⟨a,b⟩), b\big\rangle) \; \mid \; q∈ Q, a ∈ Σ, b ∈ \Pi \rbrace \\ ∪ \lbrace (⟨q, b⟩, \langle q, a\rangle) \; \mid \; q∈ Q, a ∈ Σ, b ∈ \Pi \rbrace\\ ∪ \lbrace (q_0, ⟨q_0, a⟩) \; \mid \; a ∈ Σ\rbrace$

Priorities:

$Ω^G(q, c) = Ω^𝒜(q) \qquad ∀ q ∈ Q, c ∈ Σ ∪ \Pi$
2. Finite parity game:

Solve $G, q_0$: memoryless winning strategy

• If P1 has a winning strategy: $τ: q_0 ⟼ τ(q_0) = x ∈ Σ$. For every $τ: (q, b) ⟼ τ(q,b) ∈ Σ$

• If P0 has a winning strategy $σ: Q×Σ ⟶ \Pi$, then $ν(q, a) = δ(q, (a, σ(q,a)))\\ β(q,a) = σ(q,a) \qquad ∀ q ∈ Q, a∈ Σ$

Then every execution $(\overline{x}, \overline{y}) ∈ (Σ × \Pi)^ω$ that follows $ℳ$ is accepted by $𝒜$

Tags:

Updated: