Professional Documents
Culture Documents
1 s2.0 S1474667015374048 Main
1 s2.0 S1474667015374048 Main
1 s2.0 S1474667015374048 Main
Keywords: Timed supervisory control, partial observation, relative observability, relative weak
observability, timed discrete-event systems
Q
normality {Tσ |σ ∈ Σact } is the set of marker states. Starting from
q0 , TDES G executes state transitions in accordance with
its transition function δ. Let q = (a, {tα |α ∈ Σact }) ∈ Q
relative observability relative weak observability and σ ∈ Σact ; δ is defined at (q, σ), written δ(q, σ)!, if δact
of Gact is defined at (a, σ) (i.e. δact (a, σ)!) and timer tσ
satisfies (i) 0 ≤ tσ ≤ uσ − lσ when σ ∈ Σspe , and (ii)
tσ = 0 when σ ∈ Σrem . The new state q 0 = δ(q, σ) is given
observability weak observability by q 0 = (δact (a, σ), {t0α |α ∈ Σact }), where t0σ is set to be
(Lin and Wonham, 1995) (Takai and Ushio, 2006) its default value tσ0 as in (2); for other timers tα , α 6= σ,
the reader is referred to detailed updating rules given in
Fig. 1. Observability concepts and their relations in super- Brandin and Wonham (1994); Wonham (2013),Chapter 9.
visory control of timed DES under partial observation. On the other hand, δ(q, tick)! if no timer of a prospective
event is zero, and q 0 = δ(q, tick) = (a, {t0α |α ∈ Σact }), i.e.
2. PRELIMINARIES ON TIMED DES there is no change in the activity component a of q, while
the rules for updating timers are again referred to Brandin
This section reviews the timed DES (TDES) model pro- and Wonham (1994); Wonham (2013),Chapter 9.
posed by Brandin and Wonham (Brandin and Wonham Let Σ∗ be the set of all finite strings of elements in Σ =
(1994); Wonham (2013),Chapter 9). First consider the ˙
Σact ∪{tick}, including the empty string ². We introduce
untimed DES model the languages generated by TDES G in (4). The transition
Gact = (A, Σact , δact , a0 , Am ). (1) function δ is extended to δ : Q × Σ∗ → Q in the usual way.
Here A is the finite set of activities, Σact is the finite set The closed behavior of G is the language L(G) := {s ∈
of events, δact : A × Σact → A is the (partial) activity Σ∗ |δ(q0 , s)!}, and the marked behavior is Lm (G) := {s ∈
transition function, a0 ∈ A is the initial activity, and L(G)|δ(q0 , s) ∈ Qm }. Let K ⊆ Σ∗ be a language; its prefix
Am ⊆ A is the set of marker activities. Let N denote closure is K := {s ∈ Σ∗ |(∃t ∈ Σ∗ )st ∈ K}. We say K is
the natural numbers {0, 1, 2, ...}. We introduce time into Lm (G)-closed if
Gact by assigning to each event σ ∈ Σact a lower time K ∩ Lm (G) = K. (5)
bound lσ ∈ N and an upper time bound uσ ∈ N ∪ {∞},
such that lσ ≤ uσ ; typically, lσ represents a delay in TDES G is nonblocking if Lm (G) = L(G).
communication or in control enforcement, while uσ is often To use TDES G in (4) for supervisory control, it is neces-
a hard deadline imposed by legal specification or physical sary to specify certain transitions that can be controlled
necessity. With these assigned time bounds, the event set by an external supervisor. First, as in the untimed theory
˙ rem (∪˙
Σact is partitioned into two subsets: Σact = Σspe ∪Σ Wonham (2013), we need a subset of events that may be
denotes disjoint union) with Σspe := {σ ∈ Σact |uσ ∈ N} disabled. Since disabling an event usually requires prevent-
and Σrem := {σ ∈ Σact |uσ = ∞}; here “spe” denotes ing that event indefinitely from occurring, only remote
“prospective”, i.e. σ will occur within some prospective events belong to this category. Thus let a new subset
time (with a finite upper bound), while “rem” denotes Σhib ⊆ Σrem denote the prohibitible events; the supervisor
“remote”, i.e. σ will occur at some indefinite time (with is allowed to disable any prohibitible event. Next, and
no upper bound), or possibly will never occur at all. specific to TDES, we bring in another category of events
A distinguished event, written tick, is introduced which which can preempt event tick. Note that tick may not be
represents “tick of the global clock”. Attach to each event disabled, inasmuch as no control technology can stop the
σ ∈ Σact a (countdown) timer tσ ∈ N, whose default value global clock indefinitely. On this basis let a new subset
tσ0 is set to be Σf or ⊆ Σact denote the forcible events; a forcible event
½ is one that preempts event tick: if, at a state q of G,
uσ if σ ∈ Σspe ,
tσ0 := (2) tick is defined and so are one or more forcible events,
lσ if σ ∈ Σrem .
then tick can be effectively erased from the current list
When timer tσ > 0, it decreases by 1 (counting down) if of defined events (contrast with indefinite erasure). There
event tick occurs; and when tσ = 0, event σ must occur is no particular relation postulated a priori between Σf or
(resp. may occur) if σ ∈ Σspe (resp. if σ ∈ Σrem ). Note and any of Σhib , Σrem or Σspe ; in particular, a remote event
that while tick is a global event, each timer tσ is local may be both forcible and prohibitible. It is now convenient
(with respect to σ). Also define the (integer) timer interval to define the controllable event set Σc := Σhib ∪˙ {tick}.
Tσ by ½ Here designating both Σhib and tick controllable is to
[0, uσ ] if σ ∈ Σspe , simplify terminology. We emphasize that events in Σhib
Tσ := (3)
[0, lσ ] if σ ∈ Σrem . can be disabled indefinitely, while tick may be preempted
only by events in Σf or . The uncontrollable event set Σu is
Based on (1)-(3), the TDES model G is given by Σu := Σ − Σc = Σspe ∪(Σ˙ rem − Σhib ).
G := (Q, Σ, δ, q0 , Qm ), (4) We introduce the notion of controllability as follows. Let
Q Q
where Q := A × {Tσ |σ ∈ Σact } ( denotes Cartesian K ⊆ L(G) and s ∈ K; define the eligible event subset
product) is the finite set of states, a state q ∈ Q being of
EK (s) := {σ ∈ Σ|sσ ∈ K}. (6)
the form q = (a, {tσ |σ ∈ Σact }) (i.e. a (1 + |Σact |)-tuple);
˙
Σ := Σact ∪{tick} is the finite set of events; δ : Q × Σ → Q We say K is controllable wrt. G in (4) if, for all s ∈ K,
is the (partial) state transition function; q0 = (a0 , {tσ0 |σ ∈
Σact }) (tσ0 as in (2)) is the initial state; and Qm ⊆ Am ×
209
WODES 2014
Cachan, France. May 14-16, 2014
E ˙
(s) ∩ (Σu ∪{tick}) admissible supervisor V such that Lm (V /G) = K if and
L(G)
if EK (s) ∩ Σf or = ∅, only if
EK (s) ⊇ (7)
EL(G) (s) ∩ Σu
(i) K is observable (in (9));
if EK (s) ∩ Σf or 6= ∅.
Thus K controllable means that an event σ is eligible (ii) K is controllable (in (7));
to occur in K if (i) σ is currently eligible in L(G), and (iii) K is Lm (G)-closed (in (5)).
(ii) either σ is uncontrollable or σ = tick when there is
no forcible event currently eligible in K. Controllability While controllability and Lm (G)-closedness are properties
plays the central role in the TDES supervisory control closed under set union, observability is not; consequently
framework developed in Brandin and Wonham (1994) for the supremal sublanguage that satisfies the above three
the case of full-event observation. conditions (or the optimal supervisor) need not exist in
general. This problem motivates us to propose the concept
3. SUPERVISORY CONTROL OF TDES WITH of relative observability, which is closed under set union.
PARTIAL OBSERVATION
3.2 Relative Observability of TDES
The Brandin-Wonham TDES framework has been ex-
tended to the partial-event observation case by Lin and Fixing a sublanguage C ⊆ Lm (G), we introduce relative
Wonham (1995), where the concepts of observability and observability which sets C ⊆ L(G) to be the ambient
normality of TDES were introduced. This work is first language in which observability is tested.
reviewed. Then we introduce relative observability, which Definition 2. Let K ⊆ C ⊆ Lm (G). We say K is relatively
is weaker than normality, stronger than observability, and observable with respect to C, G, and P , or simply C-
closed under set union. Computation of the supremal rel- observable, if for every pair of strings s, s0 ∈ Σ∗ with
atively observable (and controllable, Lm (G)-closed) sub- P s = P s0 there holds
language of a given language will be discussed.
(∀σ ∈ Σc ) sσ ∈ K, s0 ∈ C, s0 σ ∈ L(G) ⇒ s0 σ ∈ K (10)
3.1 Lin and Wonham’s Result ˙
where Σc = Σhib ∪{tick}.
Let Σo ⊆ Σ be a subset of observable events. Define the Relative observability was first proposed in Cai et al.
natural projection P : Σ∗ → Σ∗o according to (2013) for untimed DES. For TDES here, we extend the
concept by accounting for the event tick which may be
P (²) = ², ² is the empty string; preempted only by a forcible event, in contrast with direct
½
², if σ ∈
/ Σo , disablement of prohibitible events.
P (σ) = (8)
σ, if σ ∈ Σo ;
If C1 ⊆ C2 ⊆ L(G) are two ambient languages, it follows
P (sσ) = P (s)P (σ), s ∈ Σ∗ , σ ∈ Σ. easily from Definition 2 that C2 -observability implies C1 -
As usual, P is extended to P : P wr(Σ∗ ) → P wr(Σ∗o ), observability. Namely, the smaller the ambient language,
where P wr(·) denotes powerset. Write P −1 : P wr(Σ∗o ) → the weaker the relative observability. In the special case
P wr(Σ∗ ) for the inverse-image function of P . where the ambient C = K, Definition 2 becomes standard
observability (Lin and Wonham (1995)) for the given K.
A supervisor V under partial observation is any map This immediately implies the following.
V : P L(G) → P wr(Σ). Denote by V /G the closed-loop
system where G is under the supervision of V ; then the Proposition 3. If K ⊆ C is C-observable, then K is also
closed language L(V /G) ⊆ L(G) is defined inductively observable.
according to
The reverse statement need not be true (Cai et al. (2013)).
(i) ² ∈ L(V /G); An important way in which relative observability differs
from observability is the exploitation of a fixed ambient
(ii) s ∈ L(V /G), σ ∈ V (P s), sσ ∈ L(G) ⇒ sσ ∈ L(V /G). C ⊆ L(G). Let Ki ⊆ C, i = 1, 2. For (standard)
The marked language Lm (V /G) of V /G is defined by observability of each Ki , one checks lookalike string pairs
Lm (V /G) := L(V /G) ∩ Lm (G). only in Ki , ignoring all candidates permitted by the other
language. Observability of Ki is in this sense ‘myopic’, and
A supervisor V is nonblocking if Lm (V /G) = L(V /G), consequently, both Ki being observable need not imply
and admissible if for each s ∈ L(V /G), that their union K1 ∪ K2 is observable. The fixed ambient
language C, by contrast, provides a ‘global reference’:
(i) Σu ⊆ V (P s);
no matter which Ki one checks for relative observability,
(ii) EL(G) (s) ∩ V (P s) ∩ Σf or = ∅ & tick ∈ EL(G) (s) ⇒ all lookalike string pairs in C must be considered. This
tick ∈ V (P s). more stringent requirement renders relative observability
algebraically well-behaved: an arbitrary union of relatively
˙
Let K ⊆ Lm (G), and recall Σc = Σhib ∪{tick}. We say K observable languages is again relatively observable.
is observable (wrt. G and P ) if for every pair of strings
s, s0 ∈ Σ∗ with P s = P s0 there holds Proposition 4. Let Ki ⊆ S C, i ∈ I (some index set), be C-
observable. Then K = {Ki | i ∈ I} is also C-observable.
(∀σ ∈ Σc )sσ ∈ K, s0 ∈ K, s0 σ ∈ L(G) ⇒ s0 σ ∈ K. (9)
Theorem 1. (Lin and Wonham (1995)) Let K ⊆ Lm (G) A proof is in Cai et al. (2013) (identical to the untimed
be a nonempty language. There exists a nonblocking, case). Whether or not K ⊆ C is C-observable, write
210
WODES 2014
Cachan, France. May 14-16, 2014
211
WODES 2014
Cachan, France. May 14-16, 2014
Theorem 6. (Takai and Ushio (2006)) Let K ⊆ Lm (G) (1) K is C-observable with respect to Σhib , i.e. for every
be a nonempty language. There exists a nonblocking pair of strings s, s0 ∈ Σ∗ with P s = P s0 there holds
supervisor V such that Lm (V /G) = K if and only if (∀σ ∈ Σhib ) sσ ∈ K, s0 ∈ C, s0 σ ∈ L(G) ⇒ s0 σ ∈ K.
(i) K is weakly observable (in (14));
(2) For every pair of strings s, s0 ∈ Σ∗ with P s = P s0 there
(ii) K is controllable (in (7)); holds
(iii) K is Lm (G)-closed (in (5)). s.tick ∈ K, s0 ∈ C, s0 .tick ∈ L(G), s ≡ s0 ⇒ s0 .tick ∈ K.
Like observability, weak observability is not closed under The first condition above is the relative observability of
set union; consequently the supremal sublanguage that K with respect to Σhib . The second condition deals with
satisfies the above three conditions (or the optimal su- the event tick: two lookalike strings s, s0 ∈ C which satisfy
pervisor) need not exist in general. This motivates us to s ≡ s0 are required to have identical one-step continuations
propose relative weak observability below, which is closed of tick, if allowed in L(G), with respect to membership in
under set union. K. This is weaker than relative observability with respect
to tick, inasmuch as the requirement is imposed only on
Remark 1. The implementation of the supervisor V =
lookalike strings satisfying s ≡ s0 . Therefore the following
(V1 , V2 ) in Theorem 6 is as follows. After a string s ∈ L(G)
result is immediate.
such that s.tick ∈ L(G), V observes the string t = P s ∈
P L(G). Then V enables all events in V1 (t), and forces Proposition 8. If K ⊆ C is C-observable, then K is also
all events in V2 (t) = F (t). If one or more events in F (t) weakly C-observable.
is eligible after s, then tick is preempted; if no event in
F (t) is eligible after s, then tick is enabled. In comparison, As a corollary of Propositions 8 and 5, relative weak
the implementation of the supervisor V in Theorem 1 is observability is weaker than normality. Next, we show
simpler inasmuch as no explicit F (t) is needed for tick that relative weak observability is stronger than weak
preemption; indeed, V directly decides to enable or disable observability.
tick, and controllability ensures the availability of forcible Proposition 9. If K ⊆ C is weakly C-observable and
events for the disabling/preempting action. controllable, then K is also weakly observable.
For the proof refer to Cai et al. (2014). As with relative
4.2 Relative Weak Observability observability, the fixed ambient language C, as well as the
equivalence relation ≡, renders relative weak observability
Fixing a sublanguage C ⊆ Lm (G), we introduce rela- algebraically well-behaved: an arbitrary union of relatively
tive weak observability which sets C ⊆ L(G) to be the weakly observable languages is again relatively weakly
ambient language (as is done in Definition 2 for relative observable.
observability). The key idea here is to distinguish differ- Proposition 10. Let Ki ⊆ C, i ∈ IS (some index set), be
ent ‘control patterns’ for tick preemption in each set of
weakly C-observable. Then K = {Ki | i ∈ I} is also
lookalike strings; we do so by imposing on each such set
a special equivalence relation. The equivalence classes of weakly C-observable.
this equivalence relation have mutually disjoint subsets
For the proof refer to Cai et al. (2014). Whether or not
of forcible events, so that in each equivalence class tick
preemption may be carried out independently. K ⊆ C is weakly C-observable, write
WO(K, C) := {K 0 ⊆ K | K 0 is weakly C-observable}
Let P : Σ∗ → Σ∗o and s ∈ L(G). Write [s] := {s0 ∈ (15)
L(G)|P s0 = P s} for the set of lookalike strings to s in
L(G). Define a binary relation ≡ on [s] as follows: for for the family of weakly C-observable sublanguages of
all s, s0 ∈ [s], s ≡ s0 if either (i) EL(G) (s) ∩ EL(G) (s0 ) ∩ K. Note that the empty language ∅ is trivially weakly
Σf or 6= ∅ or (ii) there exist s1 , ..., sk ∈ [s], k ≥ 1, such that C-observable, thus a member of WO(K, C). By Propo-
sition 10, moreover, WO(K, C) has a unique supremal
EL(G) (s) ∩ EL(G) (s1 ) ∩ Σf or 6= ∅ element supWO(K, C) given by
.. [
. supWO(K, C) := {K 0 | K 0 ∈ WO(K, C)}. (16)
EL(G) (sk ) ∩ EL(G) (s0 ) ∩ Σf or 6= ∅. This is the supremal weakly C-observable sublanguage of
In words, two strings s, s0 ∈ [s] satisfy s ≡ s0 if either (i) K. The next subsection presents an algorithm to compute
they are followed by some common forcible events that are supWO(K, C).
eligible in L(G), or (ii) there is a finite chain of strings in
[s] that ‘connects’ s to s0 through some common forcible 4.3 Algorithm
events that are eligible in L(G). This implies that s ≡ s0
is false (s, s0 ∈ [s]) if and only if for every s00 ∈ [s] with As noted immediately above Proposition 8, the only dif-
s00 ≡ s0 there holds EL(G) (s) ∩ EL(G) (s00 ) ∩ Σf or = ∅. ference between relative weak observability and relative
It is easily verified that ≡ is reflexive, symmetric, and observability is the treatment of the event tick: in the
transitive, and thus an equivalence relation on [s]. former, essentially, tick may be treated independently for
Definition 7. Let K ⊆ C ⊆ Lm (G). We say K is relatively lookalike strings that do not belong to the same equiva-
weakly observable with respect to C, G, and P , or simply lence class of ≡. Thus our idea of computing the supremal
weakly C-observable, if the following two conditions hold: relatively weakly observable sublanguage of a language K
212
WODES 2014
Cachan, France. May 14-16, 2014
is as follows: (1) identify equivalence classes of ≡, and ing behavior improvement; in practice the latter will be
relabel tick using distinct event labels tick1 , tick2 , ... for case-dependent. Nevertheless, since we have designed algo-
distinct equivalent classes; (2) apply the algorithm in Cai rithms for both observability concepts, our suggestion is as
et al. (2013) to compute the supremal relatively observable follows. First compute the supremal relatively observable
sublanguage of K; and finally (3) relabel tick1 , tick2 , ... sublanguage K1 (of a given specification language K);
back to tick. if the tick preemption behavior of K1 is ‘satisfactory’,
then use K1 . Otherwise, compute the supremal relatively
Let G = (Q, Σ, δ, q0 , Qm ), C, and K be finite-state (trim) weakly observable sublanguage K2 of K; comparing K2
TDES (as in (4)) with marked languages Lm (G), C, and with K1 , if the improvement of tick preemption behavior
K, respectively. is ‘significant’, then use K2 .
Algorithm 2: Input G, C, K, and P : Σ∗ → Σ∗o .
5. CONCLUSIONS
1. For each t ∈ P (L(G)), use the subset construction
technique (Wonham (2013), Section 2.5) to find the subset
We have studied two new observability concepts in su-
Q(t) := {q ∈ Q|(∃s ∈ P −1 (t))δ(q0 , s) = q}. pervisory control of timed DES under partial observa-
For each q ∈ Q(t), write EL(G) (q) := {σ ∈ Σ|δ(q, σ)!}. tion. First, we have introduced timed relative observabil-
Then for each pair (q, q 0 ) ∈ Q(t) × Q(t), q ≡ q 0 if either ity, which is stronger than timed observability, weaker
(i) EL(G) (q) ∩ EL(G) (q 0 ) ∩ Σf or 6= ∅ or (ii) there exist than normality, and closed under set union. An algorithm
q1 , ..., qk ∈ Q(t), k ≥ 1, such that has been presented for computing the supremal relatively
observable, controllable, and Lm (G)-closed sublanguage
EL(G) (q) ∩ EL(G) (q1 ) ∩ Σf or 6= ∅ of a given language. Second, we have identified timed
.. relative weak observability, which is stronger than weak
. observability, weaker than normality, and closed under set
0
EL(G) (qk ) ∩ EL(G) (q ) ∩ Σf or 6= ∅. union. An algorithm has been presented for computing
Thus for each Q(t) we identify the equivalence classes of the supremal relatively weakly observable sublanguage of
≡, say Q1 (t), Q2 (t), .... For tick defined at some state in a given language.
Qi (t), i = 1, 2, ..., relabel it by ticki . Do the corresponding
relabeling in C and K, and denote the relabeled generators REFERENCES
by G0 , C0 , and K0 . Brandin, B. and Wonham, W. (1994). Supervisory control
2. Apply the algorithm in Cai et al. (2013) with inputs of timed discrete-event systems. IEEE Trans. Autom.
G0 , C0 , and K0 , to compute K0↑ , where Lm (K0↑ ) is the Control, 39(2), 329–342.
supremal C-observable sublanguage of K. Cai, K., Zhang, R., and Wonham, W.M. (2013). On
relative observability of discrete-event systems. In Proc.
3. Relabel the events ticki in K0↑ by tick, and denote the 52nd IEEE Conf. on Decision and Control, 7285–7290.
result by K↑ . Lm (K↑ ) is the supremal weakly C-observable Florence, Italy.
sublanguage of K. Cai, K., Zhang, R., and Wonham, W.M. (2014).
Step 1 of Algorithm 2 above is in the worst case exponen- On relative observability of timed discrete-event
tial in the state size of G; this is the overhead complexity systems. Technical report, Systems Control
as compared to Algorithm 1. For an illustrative example Group, ECE Dept, University of Toronto. Online
of Algorithm 2, the reader is referred to Cai et al. (2014). at https://sites.google.com/site/kaikai627/publication.
Cieslak, R., Desclaux, C., Fawaz, A.S., and Varaiya, P.
Finally, for a nonempty specification language K ⊆ (1988). Supervisory control of discrete-event processes
Lm (G), we modify Algorithm 1 by using Algorithm 2 in with partial observations. IEEE Trans. Autom. Control,
Step 3 to compute the supremal weakly K-observability, 33(3), 249–260.
controllability, and Lm (G)-closed sublanguage of K. De- Lin, F. and Wonham, W.M. (1988). On observability of
note this supremal sublanguage by K ↑ ; by Proposition 9, discrete-event systems. Inform. Sci., 44, 173–198.
K ↑ is weakly observable, controllable, and Lm (G)-closed. Lin, F. and Wonham, W.M. (1995). Supervisory control of
Therefore, by Theorem 6, there exists a nonblocking su- timed discrete-event systems under partial observation.
pervisor V such that Lm (V /G) = K ↑ . IEEE Trans. Autom. Control, 40(3), 558–562.
Takai, S. and Ushio, T. (2006). A new class of supervisors
Remark 2 (tradeoff between timed relative observability for timed discrete event systems under partial observa-
and relative weak observability). We have derived two ob- tion. Discrete Event Dynamic Systems, 16(2), 257–278.
servability concepts for timed supervisory control. Timed Wonham, W.M. (2013). Supervisory Control of Discrete-
relative observability is conceptually simpler (since its Event Systems. Systems Control Group, ECE Dept,
requirement is imposed only on lookalike strings), allows University of Toronto, updated July 1, 2013. Available
easier implementation (see Remark 1), but the resulting online at http://www.control.toronto.edu/DES.
tick preemption behavior is generally more restrictive. On Wonham, W.M. and Ramadge, P.J. (1987). On the
the other hand, timed relative weak observability requires supremal controllable sublanguage of a given language.
extra information about the equivalence relation ≡ on SIAM J. of Control and Optimization, 25(3), 637–659.
lookalike strings, to trade for generally more permissive
tick preemption behavior. The decision as to which ob-
servability concept to use therefore depends on how much
extra information is needed to achieve the correspond-
213