6. CONTROL DE VIBRACIONES
6.6. Absorbedores dinámicos de vibraciones
winning iff, for all pairs(RG, GG)∈ Ω GP,
η|= ♦RG → ♦GG,
i.e., if there are infinitely many vertices in the play that visit RG then there have to be infinitely many vertices that visitGG.
Infinite plays that have an infinite observation, i.e. that visit vertices inVP
ctrandVenvP in alter-
nation, are only winning ifσ is accepted byP, i.e., σ ∈ L(P), which is ensured by ΩPand
if the observation isFair[A]-schedulable, which is ensured by Ωfair. The infinite plays that
have finite observations, i.e., that reach a vertexv∈ VP
termand stay there forever due to the
self-loop edge, are winning iffv /∈ Rterm, as otherwiseRterm is visited infinitely often, but
Gterm, being empty, is never visited. This is the case for all fail(s) vertices of the underlying
game, as these always represent failure to ensure admissibility. This is likewise the case for those vertices where the corresponding state inP disallows divergence or termination, respectively.
Lemma 7.3. Let η = v0 ,→ v1 ,→ . . . ,→ vt ,→ vt ,→ . . . be an infinite play in GP
reaching a vertexvt= (v, p)∈ VtermP . Then,η is winning iff the three following conditions
are satisfied:
(1) v 6= fail(s) for some state s,
(2) v = div(s) for some state s implies p∈ Accdiv
(3) v = stop(s) for some state s implies p∈ Accstop
Remark 7.4. The Streett pair(Rterm, Gterm) is the only pair in Ω
GP that contains vertices
fromVP
term and likewise does not contain vertices not in VtermP . Plays that visits a vertex
inVP
term then stays infinitely in that vertex. The pair (Rterm, Gterm) is thus the only pair
that determines whether these plays are accepted or not and furthermore does not affect the acceptance of plays that never visitVP
term. It is therefore possible to avoid adding an
extra pair(Rterm, Gterm) by combining it with one or more of the other Streett pairs, without
changing the set of accepted plays in the game.
7.3
Constructing most general controllers for Streett objectives
We will now show how to construct a most general controller for the objective given byP from the solution of theω-regular gameGP, i.e., the set of winning vertices for player 1
and a deterministic finite memory winning strategy in the gameGP from all the winning
vertices. It is well known that two-player, perfect information turn-basedω-regular games are determined [GTW02], i.e., that each vertex in the game is winning for exactly one of the players, that there are deterministic finite-memory winning strategies for player 1 from each of his winning vertices and that both the winning vertices and winning strategies can be computed.
A deterministic winning strategy in the complete-information Streett game correspond to an admissible decision function (cf. Lemma 5.10). Given the set of winning vertices and a family of winning strategies inGP, i.e., one decision function from each winning vertex, we can construct a most general strategy as follows. The strategy operates in two phases. In the
7.3. Constructing most general controllers for Streett objectives Chapter 7. Most general controllers for ω-regular objectives
first phase, the strategy allows all legal choices that allow staying in the region of winning vertices. The fairness condition of the strategy is used to detect violations of the Streett winning condition in the long run and forces the strategy into its second phase. There, it plays according to the decision function (deterministic winning strategy in GP) for the game vertex where the switch to the second phase occurred.
As we are interested in constructing controllers, we rely on a family of finite-memory win- ning strategies for each vertex. In practice, such finite-memory winning strategies share a common set of modes that correspond to the vertices of the game plus some additional memory. If that is not the case, i.e., each finite-memory winning strategy has its own, sep- arate set of modes, it is easy to construct a “global” winning strategy by taking the disjoint union of the different modes. We can thus consider the winning strategies from the different vertices as a simplified controller, without annotations or fairness condition and having a de- cision function template that offers exactly a single choice. In addition, we need a function mapping the winning vertices to the corresponding mode.
Definition 7.5. LetWGP be the set of vertices inGP that are winning for player 1. A finite-
memory game strategyw = (Mw, ∆w, µw, initw) for player 1 consists of
• Mw, a finite set of modes,
• initw : WGP ∩ V P
ctr → Mw, a function determining the initial mode for a given
winning vertex,
• ∆w: Mw→ 2Obs, a function providing the choice of the successor vertex, and
• µw : Mw× Obs → Mw, a function updating the mode for the observable chosen by
player 2.
Letη = v0 ,→ v00 ,→ v1,→ v01,→ . . . be a play in GP, starting from a vertexv0 ∈ W(GP)∩
VP
ctrand visiting only vertices fromVctrP ∪ VenvP . Letσ = β1β2β3. . . be the observation of
η and let m0, m1, m2, . . . be the corresponding sequence of modes, with m0 = initw(v0)
andmi+1 = µw(mi, βi). We say the play η is played according to w iff allVenvP vertices
vi0 =hSi, Oii in η satisfy Oi = ∆w(mi). A play that eventually visits onlyVtermP -vertices
is played according tow iff its maximal prefix consisting entirely ofVctrP ∪ VenvP -vertices is played according tow.
A finite-memory game strategyw is winning if all maximal plays played according to w and starting in a vertex inWGP are winning inGP. In particular, all vertices visited in plays
played according tow are members ofWGP.
We now construct a most general controller for objectiveP from a solution of the complete- information gameGP, i.e., a set of verticesWGP ⊆ V
Pwhere player 1 can ensure winning
inGP and a winning finite-memory game strategyw = (Mw, ∆w, µw, initw) in case that
the initial vertex([Q0]∗, p0)∈ WGP is winning for player 1.
The controller consists of two phases. In the first phase, it ensures that the winning region WGP will not be left. The fairness condition ensures that non-winning plays eventually
switch to the second phase, where the controller then plays according tow. In case that the initial vertex of the game is winning, i.e.,([Q0]∗, p0)∈ WGP, we construct a controller
Chapter 7. Most general controllers for ω-regular objectives 7.3. Constructing most general controllers for Streett objectives
• a set Ann = {1, 2} for the two phases,
• the set of modes M = M1 ˙∪ Mw, which is the disjoint union of the modes of the first
phaseM1 =WGP∩ V P
ctrand of the second phaseMw,
• the initial mode m0= ([Q0]∗, p0)∈ M1 being the initial vertex of the game,
• the decision function template ∆ : M → 22Obs
with ∆(m) =
(
hO, 1i, hO, 2i : O ∈ ∆1(m)
form∈ M1 hO, 2i : O = ∆w(m) form∈ Mw,
where hO, ai denotes the set {hβ, ai : β ∈ O} and ∆1 : M1 → 22 Obs
is the deci- sion function template that provides all the choices that ensure staying in the winning region, i.e.,
∆1( (S, p) ) =
O : (S, p) ,→P (hS, Oi, p) and (hS, Oi, p) ∈ WGP
, • the next-mode function µ : M × Obs → M, with
µ(m,hβ, ai) = µ1(m, β) form∈ M1anda = 1 initw µ1(m, β) ∈ MG form∈ M1anda = 2 µw(m, β) form∈ MGanda = 2
and whereµ1 is the function that returns the unique β-successor of the vertex m in
GP as in the previous chapter.
• the fairness condition fair = (F1, . . . , Fk) for the Streett winning condition
ΩGP ={(R G 1, G G 1), . . . , (R G k, G G
k)} of GP such that, for16 i 6 k:
Fi = {(m, hO, 2i) : m∈ M1∩ RGi andhO, 2i ∈ ∆(m)}
∪ {(m, hO, 1i, (m, hO, 2i) : m = (S, p) ∈ M1, O∈ ∆1(m) and
(hS, Oi, p) ∈ GGi orm∈ G G i}.
The controller starts in the first phase, in the mode corresponding to the initial game vertex. In the first phase, it offers choices according to∆1, which ensures staying in the winning
region and is offered in two variants: one annotated with 1, the other annotated with 2. The controller stays in the first phase until an annotated observable with annotation2 is observed. In this case, the controller switches to phase two, the initial mode in the game strategyw for the corresponding vertex. In this second phase, the annotations become irrel- evant and are fixed arbitrarily at annotation2. Now the controller behaves according to ∆w
andµw, i.e., the game strategyw.
The fairness condition of the controller ensures that the controller may only stay in the first phase as long as the corresponding play satisfies the Streett winning condition in the game: If there is a Streett pair(RGi, GGi) where the mode (game vertex) infinitely often only occurs inRGi but not inGGi, the fairness condition requires that eventually a choice annotated with annotation2 is chosen, switching to the second phase. On the other hand, if for every Streett pair(RGi, GGi) where RGi is visited infinitely often in the first phaseGGi is visited infinitely often as well, it is fine to always schedule choices with annotation1, and the controller can stay in the first phase.
7.3. Constructing most general controllers for Streett objectives Chapter 7. Most general controllers for ω-regular objectives
Theorem 7.6 (Soundness of the constructed controller forP).
(a) If([Q0]∗, p0) ∈ WGP then the strategySC induced by the controllerC constructed
above enforcesP.
(b) If there exists an admissible decision functiond that enforcesP then ([Q0]∗, p0)∈ WGP andd is an instance of SC.
Hence, if the objective specified byP is enforceable then SC is a most general strategy
enforcingP. Proof.
ad (a). To show that SC enforces P, we have to show that SC is admissible and that
π ∈ LA(P) for all initial SC-pathsπ, i.e., that all instances d of SCare admissible and that
for all instancesd all d-paths are contained inLA(P).
We will first relate SC-paths with the executions in the controller and the corresponding
plays inGP.
Letπ be an initial SC-path and letπCbe one of the corresponding initialC ./A-paths (see
Lemma 4.2) withπC|A = π. Let
π0= m0 β1,a1 −−−→ m1 β2,a2 −−−→ m2 β3,a3 −−−→ . . . ∈ M × (Obs × M)∞
be the corresponding execution inC, i.e., with invisible actions removed, the states projected on the modes and with the annotated observables labeling the transitions.
Letd be a corresponding decision function and let O01O20O03. . . ∈ (2Obs)∞ be the corre- sponding sequence of annotated controller choices of an annotated instance corresponding to d, with O0i ∈ ∆(mi) and hβi, aii ∈ Oi0 for 1 6 i 6 |π0| and satisfying the fairness
condition of the controller. This corresponds to a maximal, initial play inGP
η = v0 ,→P v00 β1 ,→P v1 ,→P v10 β2 ,→P v2 ,→P v20 β3 ,→P . . . , wherev0i = (hS, Oi, p) ∈ VP
envforvi = (S, p) and where O = Oi0|Obs ∈ ∆(mi). By con-
struction of the controller, all the vertices in the play belong to WGP. Furthermore,η is
winning forΩGP. In case that at some point the controller switches to the second phase, the
suffix ofη after the phase switch is winning because it is played according to w. But then the whole play is winning, because a play in a Streett game is winning if any suffix is winning. In the other case, where no switch to the second phase occurs, the playη is winning because it satisfies the fairness conditionfair. Assume by contradiction that η is not winning. Then there exists a Streett acceptance pair (RG, GG) ∈ ΩGP such thatη |= ♦R
G ∧ ¬♦GG.
LetF be the corresponding element of the controller’s fairness condition fair. As there is no switch of phase, for alli> 1, the choices O0iare annotated with1. As η|= ♦RG, there
are infinitely many modesmi ∈ M1∩ RG, i.e., whereF(β1, . . . , βi) 6= ∅. On the other
hand, asη |= ♦¬GG, from some indexj ∈ N on it is the case that O0
Chapter 7. Most general controllers for ω-regular objectives 7.3. Constructing most general controllers for Streett objectives
for alli > j. But this violates condition (I2) in Definition 3.2 due to the fact that d is an instance ofC. We conclude that η is winning.
We now show thatSC is admissible. Suppose by contradiction thatSC is not admissible
by violating one of the conditions (A1), (A2) or (A3) of admissibility. By Lemma 5.9, this corresponds to the fact thatη reaches a vertex of the form fail(s) for some state s. But any maximal play that reaches a vertex of the form fail(s) is not winning for winning condition ΩGP as fail(s) 6∈ WGP, which contradicts the fact that all vertices of η belong toWGP.
It remains to show thatSC respectsFair[A]. Suppose by contradiction that SC does not
respectFair[A]. Then, there exists a Fi ∈ Fair[A] and an initial SC-pathπ such that obs(π)
is notFi-schedulable. But then the corresponding playη does not satisfy the Streett winning
condition(Rfair
i , Gfairi ), which contradicts the fact that η is winning.
Similarly, we show thatπ∈ LA(P) for all initial SC-pathsπ, by showing that π∈ LA(P)
for all initiald-paths and for all instances d of SC.
We first consider the case that the pathπ and the observation σ = β1β2. . . βn are finite.
Then, the corresponding playη eventually always visits a vertex (stop(s), p) with s|A =
last(π). As the play is winning, p ∈ Accstop and therefore π ∈ LA(P). Likewise, if
π is infinite then π is divergent and the corresponding play eventually stays in a vertex (div(s), p) with p∈ Accdiv, i.e.,π diverges from the state s|Aandπ∈ LA(P).
It remains to show thatπ ∈ LA(P) in the case that both π and σ are infinite, which is the
case if observationσ is accepted byP. We show that σ is accepted by P if the play η is winning. Note that the runξ = p0p1p2, . . . for σ inP corresponds to the states pi in the
playη. Let (RjG, GGj) be one of the Streett pairs in the winning condition ΩP of the game
GP and let(Rj, Gj) be the corresponding Streett pair in the acceptance condition Acc ofP.
By construction ofΩGP, for alli> 0,
vi = (Si, pi)∈ RGj ⇐⇒ pi ∈ Rj and vi= (Si, pi)∈ GGj ⇐⇒ pi ∈ Gj.
Therefore, ifRGj is visited infinitely often in the playη then Rj is visited infinitely often in
the runξ. Also, if GGj is visited infinitely often in the playη then Gj is visited infinitely
often in the runξ. As this is the case for all Streett pairs in the acceptance condition ofP, the runξ inP for observation σ is accepted by P iff the play η is winning. As we know that η is winning, π∈ LA(P). We conclude that SCenforcesP for (A, Fair[A]).
ad (b). We show that if there exists an admissible decision functiond that enforcesP then the initial vertex ofGP is winning,([Q0]∗, p0)∈ WGP, andd is an instance of SC.
Claim 1. Let d be an admissible decision function that enforces P. Then, ([Q0]∗, p0)∈ WGP.
7.3. Constructing most general controllers for Streett objectives Chapter 7. Most general controllers for ω-regular objectives
As usual, w.l.o.g, we assume thatd only schedules schedulable observables. We show that ([Q0]∗, p0) ∈ WGP by showing the stronger claim that all verticesv ∈ V
P that are reach-
able by an initial, maximal play inGP played according to decision functiond, i.e., where
the player 1 choices are determined by d, are winning, v ∈ WGP. As the initial vertex
([Q0]∗, p0) is trivially reachable, it is then surely contained inWGP.
Assume by contradiction that there is an infinite, initial playη inGP where the successor
chosen by player 1 is determined byd depending on the history of the play with at least one of the vertices inη not contained inWGP.
Then, the playη either consists entirely of alternating VP
ctr andVenvP vertices or it reaches
some vertex fromVtermP . In the first case, the playη has the form η = v0 ,→P v00 β1 ,→P v1 ,→P v01 β2 ,→P v2 ,→P v02 β3 ,→P . . . ,
starting inv0 = ([Q0]∗, p0), with vi = (Si, pi) ∈ VctrP andvi0 = (hSi, d(β1. . . βi)i, pi) ∈
VP
envfor someSiandpiandi> 0.
In the second case, the playη has the form η = v0 ,→P v00
β1
,→P v1 ,→P v10 . . . ,→P vk ,→P vk0 ,→P vt ,→P vt ,→P . . . ,
starting inv0 = ([Q0]∗, p0), with vi = (Si, pi) ∈ VctrP andvi0 = (hSi, d(β1. . . βi)i, pi) ∈
VP
envfor someSiandpiand06 i 6 k and with vt∈ VtermP .
By the assumption, not all vertices inη are inWGP. We first consider the case that the play
reaches a vertex fromVP
termand thatvtis the first and only vertex inη that is not contained
inWGP.
By the construction of the game winning conditionΩGP,vt∈ W/ GP implies thatvt∈ R term,
as otherwise the infinite play consisting solely ofvtvertices would be winning for player 1.
It can not be the case thatvtis of the form(fail(s), p), as then d would be non-admissible
(Lemma 5.9). Ifvtis of the form(div(s), p), then Lemma 5.8 provides the existence of an
initial d-path π with observation obs(π) = β1. . . βk. As vtis not winning, p /∈ Accdiv.