Professional Documents
Culture Documents
REAL-TIME SCHEDULING OF
MANUFACTURING SYSTEMS
y
P. R. Kumar and T. I. Seidman
z
Abstra
t
We
onsider manufa
turing systems
onsisting of many ma
hines
and produ
ing many types of parts. Ea
h part-type requires pro
essing
for a spe
ied length of time at ea
h ma
hine in a pres
ribed sequen
e
of ma
hines. Ma
hines may require a set-up time when
hanging be-
tween part-types, and parts may in
ur a variable transportation delay
when moving between ma
hines. The goal is to dynami
ally s
hedule
all the ma
hines so that all the part-types are produ
ed at the desired
rates while maintaining bounded buer sizes at all ma
hines.
In this paper we study the intera
tion of two types of feedba
ks,
one
aused by \
y
les" of material
ow in non-a
y
li
manufa
turing
systems, and the other introdu
ed by the employment of
losed-loop
s
heduling algorithms. We examine the
onsequen
es of this intera
-
tion for the stability properties of the manufa
turing system in terms
of maintaining bounded buer levels.
First, we resolve a previously open problem by exhibiting the in-
stability of all \
learing poli
ies" for some non-a
y
li
manufa
turing
The resear
h of the rst author has been supported in part by the National S
i-
en
e Foundation under Grant No. ECS-88-02576, and in part by the Manufa
turing
Resear
h Center of the University of Illinois, and the resear
h of the se
ond author has
been supported in part by the Air For
e OÆ
e of S
ienti
Resear
h under Contra
t No.
AFOSR-87-0190.
y Department of Ele
tri
al and Computer Engineering and the Coordinated S
ien
e
Laboratory, University of Illinois, Urbana-Champaign. Mailing Address: University of
Illinois, C.S.L., 1101 W. Springeld Ave., Urbana, IL 61801. Tel. No. 217-333-7476.
Please address all
orresponden
e to the rst author.
z Department of Mathemati
s, University of Maryland Baltimore County, 5401 Wilkens
Avenue, Baltimore, MD 21228.
1
systems. Surprisingly, su
h instabilities
an o
ur even when all set-up
times are identi
ally zero, and they are indu
ed purely by starvation of
ma
hines. Simultaneously however, there
an exist
ertain exa
t sets
of initial
onditions for whi
h a deli
ate stability does hold; however,
it may not be robust. Se
ond, we exhibit suÆ
ient
onditions on the
system topology and pro
essing and demand rates whi
h ensure the
stability of distributed
lear-a-fra
tion poli
ies. These
onditions are
easy to verify. Third, we study general supervisory me
hanisms whi
h
will stabilize any poli
y. One su
h universal stabilization me
hanism
requires only a supervisory level whi
h trun
ates all ex
essively long
produ
tion runs of any part-type, and maintains a separate priority
queue for all part-types with large buer levels. It is easily imple-
mentable in a distributed fashion at the ma
hines, and the level of
supervisory intervention
an also be adjusted as desired.
1 INTRODUCTION
SYSTEMS
and sent to buer 4.
Stage 12. At t13 := t12 + 4 x4 (t12 ), buer 4 empties, i.e. x4 (t13 ) = 0. One
has x1 (t13 ) = x1 (t12 ) + (t13 t12 ); x2 (t13 ) = 0, and x3 (t13 ) = 0. Note that
ma
hine 1 is still set-up for buer 4 at t13 , while ma
hine 2 is still set-up for
buer 3.
A straightforward
omputation shows that t13 = t1 + ( + 1 22 ) + ,
and x1 (t13 ) = + , where
:= 4 > 1;
1 2
(4 + 1)(Æ1 + Æ2 )
:= + Æ3 (4 + 1) + Æ4 ;
(1 2 )
and
4 (Æ1 + Æ2 )
:= + 4 Æ3 + Æ4 :
1 2
The important feature to note is that starting at a
ertain time T0 at
x(T0 ) = (; 0; 0; 0), with ma
hines 1 and 2 set-up to buers 4 and 3, respe
-
tively, we return at a time T1 := T0 +(+ 1 22 ) + to x(T1 ) = ( +; 0; 0; 0)
with ma
hines 1 and 2 again set-up to buers 4 and 3, respe
tively. This
pattern of behavior therefore re
urs innitely often, and we obtain,
Tn+1 = Tn + ( + 2 )x1 (Tn ) + ;
1 2
x(Tn+1 ) = (x1 (Tn ) + ; 0; 0; 0):
Noting that > 1 due to (3), we thus obtain limn!1 x1 (Tn ) = +1, prov-
ing that the buers are not bounded. Regarding the throughput of the
manufa
turing system we note that,
System output in [T0 ; Tn ℄ x1 (Tn ) x1 (T0 )
System input in [T0 ; Tn ℄
=1
Tn T 0
! +1 < 1;
2 4
by virtue of (3). Hen
e the system does not produ
e parts at the desired
(input) rate.
Case 2: No Set-Up Times
Now we turn to the situation where Æ1 = Æ2 = Æ3 = Æ4 = 0, i.e. all the
set-up times are zero. We will assume that the initial
ondition at t1 = 0 is
as before, i.e. x(t1 ) = (; 0; 0; 0) where > 0. Due to the absen
e of set-up
times, the analysis is in fa
t a little easier than before.
Stage 1. At t2 := t1 + x1 (1t1 )1 , buer 1
lears. During the interval [t1 ; t2 ℄,
1
ma
hine 2 is pro
essing buer 2, and so x(t2 ) = (0; ( 12 1)(t2 t1); 12 (t2
t1 ); 0). After t2 , sin
e buer 4 is empty, ma
hine 1
ontinues to pro
ess buer
1, but at the redu
ed rate of 1, due to starvation.
Stage 2. At time t3 = t2 + x2 (1t2 )1 , buer 2 is
leared. Thus x(t3 ) = (0; 0; +
2
t3 t1 ; 0). Sin
e ma
hine 2 starts pro
essing buer 3 at t3 , buer 4 starts
lling up, and so ma
hine 1 starts pro
essing buer 4.
Stage 3. At time t4 = t3 + 3 x3 (t3 ), buer 3
lears. Hen
e x(t4 ) = (t4
t3 ; 0; 0; ( 13 14 )(t4 t3 )).
Stage 4. At time t5 = t4 + 4 x4 (t4 ), buer 4
lears. Thus x(t5 ) = (t5
t3 ; 0; 0; 0).
A
omputation shows that x1 (t5 ) = where := 1 42 > 1 as before.
Thus at t5 , the state of the system is a magnied version of that at time t1 .
Hen
e limn!1 x1 (t4n+1 ) = +1.
The main reason for the instability of this system is that during time
intervals su
h as [t2 ; t3 ℄, ma
hine 1 is for
ed to work at the redu
ed rate of
1 on buer 1, due to its buers being starved of input. Thus, valuable pro-
essing
apa
ity is lost. Also ma
hine 2 is for
ed to be idle during intervals
su
h as [t4 ; t5 ℄ due to starved buers. Thus the dynami
instability is purely
starvation indu
ed, and os
illatory. The os
illatory nature is
aused solely
by the non-a
y
li
nature of the manufa
turing system.
The following example exhibits two further phenomena. First, it shows
that dynami
instability
an arise even if no part-type ever revisits a ma-
hine. It also shows that one
an have stability or instability, depending on
the system initial
onditions.
Example 2 Consider the system shown in Figure 3 where there are 2 part-
types and 2 ma
hines. Both part-types have an input rate of 1. Part-
type 1 rst visits buer 1 at ma
hine 1 and then buer 2 at ma
hine 2,
while part-type 2 rst visits buer 3 at ma
hine 2 and then buer 4 at
ma
hine 1. The pro
essing times of the parts in buers 1, 2, 3 and 4 are
1 ; 2 ; 3 and 4 , respe
tively, while the times for setting-up to these buers
are Æ1 ; Æ2 ; Æ3 ; Æ4 > 0.
We will assume that the parameters above satisfy,
2 + 4 > 1; (7)
Æ1 + Æ4 Æ2 + Æ3
= =: ; (8)
1 1 4 1 2 3
Æ4 < (1 3 4 ); (9)
Æ2 < (1 1 2 ); (10)
and that the
apa
ity
ondition (1) is met, i.e.
1 + 4 < 1 and 2 + 3 < 1:
(A set of feasible parameters is (1 ; 2 ; 3 ; 4 ; Æ1 ; Æ2 ; Æ3 ; Æ4 ) = ( 41 ; 12 ; 14 ; 23 ; 301 ; 15 ; 201 ; 201 ),
for example). We will analyze this system when a s
heduling poli
y of
lear-
ing type is employed; there is only one su
h
learing poli
y.
Case 1: A Periodi
Regime
First we will show that there exists an initial
ondition for whi
h all
buers are bounded; for this demonstration we do not need the assumption
(7).
Consider a time t = 0 at whi
h buers 1 and 3 have just been
leared,
and so a set-up
ommen
es for buers 2 and 4. We will assume that x2 (0) =
x4 (0) = > 0, while x1 (0) = x3 (0) = 0, as stated.
Note that ma
hine 1
ompletes the set-up for buer 4 at time t = Æ4 ,
then
lears buer 4 at time t = Æ4 + 4 . and
ompletes the set-up for buer
1 at t = Æ4 + 4 + Æ1 .
Meanwhile, ma
hine 2
ompletes the set-up for buer 2 at t = Æ2 , then
lears buer 2 at t = Æ2 + 2 , and
ompletes the set-up for buer 3 at
t = Æ2 + 2 + Æ3 .
Sin
e Æ4 + 4 < Æ2 + 2 + Æ3 , by virtue of (9) and the se
ond equality
in (8), buer 4 is
leared before ma
hine 2 starts pro
essing buer 3, and
so one does not need to
onsider the possibility that buer 4's
learing is
delayed. Similarly, due to Æ2 + 2 < Æ4 + 4 + Æ1 by virtue of (10,8), buer
2's
learing time is indeed Æ2 + 2 .
At t = Æ4 + 4 + Æ1 , buer 1 has a
umulated Æ4 + 4 + Æ1 parts, whi
h
are
leared at time t = (Æ4 + 4 + Æ1 ) + (Æ4 + 1 4 +1Æ1 ) = . Similarly buer 3
1
is
leared at time exa
tly equal to .
Thus at t = , we return to the same system state as that at t = 0, and
so we have a periodi
steady state where all buers are bounded for all time.
It is worth noting that this syn
hronization of events depends on
ondi-
tion (8) holding exa
tly. It is a rather fragile stability, as we shall see in the
next
ase.
Case 2: An Unstable Regime
Now we will
onsider a dierent initial
ondition for the system. At time
t1 , we start with x(t1 ) = (; 0; 0; 0) with > 0 large enough, and ma
hine 1
set-up for buer 4, while ma
hine 2 is set-up for buer 3. We shall abbreviate
the details of the arguments. (Conditions (8-10) are unne
essary for this
analysis).
Stage 1. Ma
hine 1
ompletes the set-up for buer 1 at t2 = t1 + Æ1 , and
x(t2 ) = (x1 (t1 ) + Æ1 ; 0; 0; 13 (t2 t1 )). Note that ma
hine 2 is pro
essing at
the redu
ed rate of 1 at buer 3 in [t1 ; t2 ℄.
Stage 2. Ma
hine 2
ompletes the set-up for buer 2 at t3 = t2 + Æ2 . x(t3 ) =
(x1 (t2 ) ( 11 1)(t3 t2 ); 11 (t3 t2 ); (t3 t2 ); x4 (t2 )).
Stage 3. Ma
hine 1
lears buer 1 at t4 = t3 + x1 (1t3 )1 . x(t4 ) = (0; x2 (t3 ) +
1
( 11 12 )(t4 t3 ); x3 (t3 ) + (t4 t3 ); x4 (t3 )).
Stage 4. Ma
hine 1
ompletes the set-up for buer 4 at t5 = t4 + Æ4 . x(t3 ) =
((t5 t4 ); x2 (t4 ) 12 (t5 t4 ); x3 (t4 ) + (t5 t4 ); x4 (t4 )).
Stage 5. Ma
hine 1
lears buer 4 at t6 = t5 + 4 x4 (t4 ). x(t6 ) = (x1 (t5 ) +
(t6 t5 ); x2 (t5 ) 12 (t6 t5 ); x3 (t5 ) + (t6 t5 ); 0).
Stage 6. Ma
hine 1
ompletes the set-up for buer 1 at t7 = t6 + Æ1 . x(t7 ) =
(x1 (t6 ) + (t7 t6 ); x2 (t6 ) 12 (t7 t6 ); x3 (t6 ) + (t7 t6 ); 0).
Stage 7. Ma
hine 1
lears buer 1 at t8 = t7 + x1 (1t7 )1 . x(t8 ) = (0; x2 (t7 ) +
1
( 11 12 )(t8 t7 ); x3 (t7 ) + (t8 t7 ); 0). Sin
e buer 4 is empty, ma
hine 1
ontinues to work on buer 1 at the redu
ed rate of 1 after t8 .
Stage 8. Ma
hine 2
lears buer 2 at t9 = t8 + x2 (1t8 )1 . x(t9 ) = (0; 0; x3 (t8 ) +
2
(t9 t8 ); 0).
Any easy way to
ompute t9 is as follows. During the period [t1 ; t9 ℄
ma
hine 2 has pro
essed the amount in buer 1 at t1 plus the amount
(t9 t1 ) that
ame in during [t1 ; t9 ℄. Thus 12 (t9 t1 Æ1 Æ2 ) = + (t9
1
t1 ), giving t9 = t1 + +2 1(Æ11+Æ2 ) . Sin
e buer 3 was not pro
essed during
2 1
[t2 ; t9 ℄; x3 (t9 ) = t9 t2 = t9 t1 Æ1 = +2 1(Æ11+Æ2 ) Æ1 .
2
Note that at t9 , the state of the system is the mirror-image of the state
at t1 . Hen
e, by repeating the mirror image of the analysis of Stages 1-8,
we arrive at a time t17 when x(t17 ) = ( + ; 0; 0; 0) where := (1 22)(1
4
4 )
h i 1
and := Æ11 +Æ22 Æ1 + 14 (Æ3 + Æ4 ) 14 1 . Also, as at t1 , ma
hine 1 is
set-up for buer 4 and ma
hine 2 for buer 3.
Hen
e the situation is the same as at t1 , ex
ept that the level of buer
1, x1 , has been magnied. Thus lim supt x1 (t) = +1.
Thus we see that the same system
an possess both stable and unstable
modes of behavior. Moreover, for instability, it is not ne
essary for a part
to revisit a ma
hine.
4 A SUFFICIENT CONDITION FOR STABIL-
rate at whi
h parts move from bp;p;i to m is d0p;i . Lastly, if p;i 2 [ln=1Gn ,
then p;k 2 [ln=1 Gn for all 1 k i 1, sin
e G1 ; : : : ; Gl were originally
minimal, and so part-type p enters the system at a rate dp at one of the
ma
hines in [ln=1 Gn . However, sin
e all the buers of su
h ma
hines are
bounded by ,
In
ow of parts in [Tn ; Tn+1 ℄ d0 [Tn+1 Tn ℄ + ;
p;i (23)
sin
e d0p;i := dp in this
ase.
Note therefore, that in all
ases, one has an inequality su
h as (23) (where
an even be set to 0 in the rst two
ases), and so,
X
In
ow of work to m in [Tn ; Tn+1 ℄ fd0p;i[p;i + : : : + p;i+p;i 1 ℄(Tn+1 Tn )
f(p;i):p;i =m;p;i 16=m g
+ (p;i + : : : + p;i+p;i 1 )g
= 0m (Tn+1 Tn ) + ;
P
where := f(p;i):p;i =m;p;i 16=m g [p;i + : : : + p;i+p;i 1 ℄. This is a simple
generalization of the bound in (16).
Now let us examine the work done by ma
hine m in the time interval
[Tn + Æbp? ;i? ;bp?n ;i?n ; Tn+1 ℄, in whi
h period ma
hine m is a
tive. If it did
n 1 n 1
not pro
ess at the maximum rate of p?1;i? throughout this time interval,
n n
then there is a time instant s 2 (Tn ; Tn+1 ℄ at whi
h V (s) = 0, sin
e ma
hine
m must have been starved at some su
h time instant. If s = Tn+1 , then
Vm (Tn+1 ) = 0. If not, let s be the last su
h time instant. In [s; Tn+1 ℄,
ma
hine m works at the maximum rate, and so the work
ompleted is (Tn+1
s). However then the work entering in [s; Tn+1 ℄
an be at most 0m (Tn+1
s) + .
Thus, by
overing all
ases, we obtain the relation,
Vm (Tn+1 ) maxf0; Vm (Tn ) (1 0m )(Tn+1 Tn ) + Æbp?n ;i?n ;bp?n ;i?n + ; max [ (1 0m )(Tn+1
Tn sTn+1
0
maxf ; Vm (Tn ) (1 m )(Tn+1 Tn) + Æbp?n ;i?n + g;
whi
h is a simple generalization of the re
urren
e bound (17). Thus by re-
peating the earlier arguments, we obtain the ultimate boundness of Vm (Tn ).
The boundedness of Vm (t) follows by noting that Vm (t) Vm (Tn ) + Æ 0m +
for all t 2 [Tn ; Tn+1 ℄.
Continuing by indu
tion, one
an next remove fGl+1 ; : : : ; Gl+q g, and by
a repeated appli
ation of this te
hnique, the proof is
ompleted.
SORY MECHANISM
The suÆ
ient
ondition (14) for CAF poli
ies to be stable for all system
initial
onditions is a more stringent requirement than the
apa
ity
ondi-
tion (1) sin
e 0m m . Hen
e rather than restri
tively requiring (14) to
hold, we are interested in obtaining poli
ies whi
h are stable whenever (1)
holds. However, as shown in Se
tion 3, dynami
instability is a very real
possibility. Is there a simple \stabilization" me
hanism by whi
h a super-
visor
an modify any s
heduling poli
y so that it be
omes stable when only
(1) holds?
We will now show there does exist su
h a \universal" safety me
ha-
nism, whi
h
an moreover be trivially implemented by a supervisor, in a
distributed fashion, at the various ma
hines. Essentially this me
hanism
onsists simply of:
i) trun
ating all long produ
tions runs, and
ii) maintaining a separate rst-
ome rst-serve (FCFS) priority queue Qm
for large buers, at ea
h ma
hine m.
For ea
h ma
hine m, let
m be a large number satisfying,
X
m (1 m ) > max Æb0 ;b : (24)
b0 2Bm
b2Bm
Sin
e m < 1 by the
apa
ity
ondition (1), su
h a
hoi
e is
learly feasible.
Se
ond, we shall arbitrarily
hoose for ea
h buer bp;i a nonnegative number
zp;i . The pre
ise operation of the supervisor is given by the following ve
rules to be implemented in a distributed fashion at ea
h ma
hine m.
The Trun
ation Rule No pro
essing run of buer bp;i at ma
hine m is
ever allowed to
ontinue beyond
m dp p;i time units.
The Rule for Entering Qm A buer bp;i enters the tail of the priority
queue Qm if (i) bp;i is not being pro
essed or set-up, and (ii) its buer level
xp;i ex
eeds zp;i .
The Buer Sele
tion Rule If Qm 6= when a pro
essing run termi-
nates, then the buer at the head of the priority queue Qm (i.e. a
ording
to a FCFS dis
ipline) is
hosen next for pro
essing.
The Rule for Leaving Qm A buer leaves the priority queue Qm when
it is taken up for pro
essing, i.e. a set-up is
ommen
ed.
The Rule for the Pro
essing Time for a Buer from Qm If a
buer bp;i from Qm is taken up for pro
essing, then it is pro
essed for exa
tly
m dp p;i time units, unless it
lears before this amount of time has elapsed.
It should be noted that the above rules do not restri
t in any way whi
h
buer is
hosen for pro
essing if Qm is empty. Thus, ex
ept for the trun-
ation rule, the supervisory me
hanism intervenes only when buer levels
be
ome large, and is otherwise unobtrusive.
A se
ond point to note is that after
ompleting a pro
essing run, a buer
bp;i may immediately enter Qm ; this happens if its buer level at the end of
the pro
essing run is larger than zp;i.
It is worth mentioning that there is
onsiderable
exibility in
hoosing
the frequen
y of supervisory intervention. For example, if zp;i := 0, then
the supervisor simply enfor
es a First-Come-First-Serve (FCFS) dis
ipline
among the non-empty buers, trun
ating their pro
essing runs after
m dp p;i
time units. Thus the supervisor always intervenes. On the other hand, if
a poli
y is already stable, and if both zp;i and
m are
hosen large enough,
then the supervisor never intervenes. Hen
e, by
hoosing zp;i and
m , the
degree of supervisory intervention
an be adjusted.
In what follows we prove that this supervisory me
hanism guarantees
stability.
The following preliminary lemma exhibits four key
onsequen
es of the
supervisory me
hanism.
Lemma 1 Suppose buer bp;i enters Qm (where m = p;i) at a time instant
tin and has its subsequent pro
essing run
ompleted at a time t
omplete .
i) Then,
t
omplete tin
m m ; (25)
where X
m :=
m (1 m ) max Æb0 ;b : (26)
b0 2Bm
b2Bm
Note that m > 0 by (24).
Pi
ii) Let p;i (t) := j =1 p;i xp;j (t) be the ba
klog of work for ma
hine m in
the prior buers visited by part-type p. Then xp;i(tin)
m dp )
p;i(t
omplete ) p;i(tin ) m dp p;i.
iii) Suppose t0 is su
h that xp;i() max(
m dp ; zp;i ) for all 2 [tin; t0 ℄.
(t )
Then t0 tin 1 + mp;idp inp;i (
m m ).
(iii) The proof is by
ontradi
tion. Suppose not, i.e. (t0 t) > (1 +
p;i (tin )
m dp p;i )(
m m ). Let t
omplete > tin be the next time its pro
essing run
omplete tin
m m , and so
is terminated. By (i) it follows that t(1) (1)
0
omplete t . Hen
e xp;i(t
omplete ) p;i again, and so bp;i again enters
t(1) (1)
(n) (n)
n su
h intervals [tin; t(1) (2) (2)
omplete ℄; [tin ; t
omplete ℄; : : : ; [tin ; t
omplete ℄ with
k 1)
t(ink) = t(
omplete for 2 k n, at the
ommen
ement of ea
h of whi
h,
xp;i (t(ink) ) p;i. Hen
e, by (ii),
n)
p;i (t(
omplete ) p;i(tin) nmdpp;i
!
p;i (tin) p;i (tin)
< p;i (tin) d (sin
e n > )
m dp p;i m p p;i m dp p;i
< 0;
whi
h is a
ontradi
tion, sin
e the ba
klog p;i(t) is always nonnegative.
(iv) At time t sin
e xp;i(t) zp;i , if bp;i is not in Qm , then it is either
being pro
essed or set-up, by the rule for entering Qm . In any event, let
s denote the time of the end of the resulting produ
tion run. If bp;i is
being pro
essed or set-up at t, then s t maxb;b0 2Bm Æbb0 +
m dp p;i,
where the rst term on the right-hand side allows for a set-up time, and the
last term represents the maximum length of a produ
tion run, due to the
trun
ation rule. In
ase bp;i 2 Qm at t, then s t
m m by (i). Sin
e
m m > maxb;b0 2Bm Æbb0 +
m dp p;i by (26), it follows that in all
ases,
s t
m m (27)
Moreover, by the trun
ation rule, the number of parts pro
essed in a run is
less than or equal to
m dp . Hen
e xp;i (s) xp;i(t)
m dp max(
m dp ; zp;i );
and so bp;i enters Qm at time s. By (iii) it follows that
!
(s)
t0 s 1 + p;i (
m m );
m dp p;i
and from (27) we thus obtain,
!
(s)
t0 t 2 + p;i (
m m ): (28)
m dp p;i
Now note that the maximum rate at whi
h p;i
an grow is dp p;i, and so
by (27), we have
p;i (s) p;i (t) + dp p;i (
m m ):
Substituting in (28) yields the result.
Our main theorem
on
erning the stabilizing property of this supervisory
me
hanism is the following.
Theorem 2 If the above supervisory me
hanism is applied to all the ma-
hines, then all buer levels are bounded over all time for all initial system
states.
Proof: Consider any part-type p. We will prove by indu
tion on i that the
buer levels xp;i(t) are bounded for 0 t < +1.
Consider the rst ma
hine m = p;1 visited by part-type p. Note that,
as a
onsequen
e, p;1 (t) = p;1xp;1 (t).
Let
t(1) := inf ft 0jxp;1 (t) gp;1 g;
and indu
tively dene for k 1,
t(k) := inf ft > t(k) jxp;1 (t) < gp;1 g;
t(k+1) := inf ft > t(k) jxp;1 (t) = gp;1 g;
with the
onvention that inf := 1. Note that if t(k) < +1, then tk < +1
by Lemma 1(iv). Moreover, if any t(k) = +1, then xp;1 (t) is bounded (sin
e
it is Lips
hitz and has no \nite es
ape time"). So let us suppose t(k) < +1
for all k. Note also that again by the Lips
hitz property, if we
an prove
that ft(k) t(k) g is a bounded sequen
e, then xp;1(t) is bounded.
Let us
onsider k 2, and note that p;1 (t(k) ) = p;1xp;1(t(k) ) = p;1gp;1 ,
and so by Lemma 1(iv),
t(k) t(k) ap;1 +
p;1 gp;1 p;1 for all k 2:
Sin
e xp;1(t) is Lips
hitz, it follows that it is bounded.
Now suppose for indu
tion that,
xp;j (t) j for all t 0; 1 j i 1:
Consider bp;i. Dene, as earlier,
t(1) := inf ft 0jxp;i(t) gp;i g;
and indu
tively dene for k 1
t(k) := inf ft > t(k) jxp;i(t) < gp;i g
t(k+1) := inf ft > t(k) jxp;i (t) = gp;i g:
Again, if t(k) = +1 for any k, then xp;i(t) is bounded. So let us suppose
that t(k) < +1 for all k. Now note for k 2,
i
X
p;i (t(k) ) = p;ixp;j (t(k) )
j =1
i 1
X
p;i j + p;ixp;i(t(k) )
j =1
i 1
X
= p;i j + p;igp;i
j =1
6 CONCLUDING REMARKS
Referen es
[1℄ K. Baker, Introdu
tion to Sequen
ing and S
heduling. New York: John
Wiley, 1974.
[2℄ R. W. Conway, W. L. Maxwell and L. W. Miller, Theory of S
heduling.
Reading, MA: Addison-Wesley, 1967.
[3℄ S. Fren
h, Sequen
ing and S
heduling. New York: John Wiley, 1982.
[4℄ E. G. Coman, Computer and Job-Shop S
heduling Theory. New York:
John Wiley, 1976.
[5℄ S. S. Panwalker and W. Iskander, \A survey of s
heduling rules," Op-
erations Resear
h, vol. 25, pp. 45{61, January{February 1977.
[6℄ S. Graves, \A review of produ
tion s
heduling," Operations Resear
h,
vol. 29, pp. 646{675, July-August 1981.
[7℄ E. L. Lawler, J. K. Lenstra and A. H. G. Rinooy Kan, \Re
ent develop-
ments in deterministi
sequen
ing and s
heduling," Deterministi
and
Sto
hasti
S
heduling, 1982. Reidel, Dordre
ht.
[8℄ J. Lenstra and A. H. G. Rinooy Kan, \Sequen
ing and s
heduling,"
Preprint.
[9℄ J. Adams, E. Balas and D. Zawa
k, \The shifting bottlene
k pro
edure
for job shop s
heduling," Management S
ien
e, vol. 34, pp. 391{401,
Mar
h 1988.
[10℄ J. F. Muth and G. L. Thompson, Industrial S
heduling. Englewood
Clis, NJ: Prenti
e Hall, 1963.
[11℄ J. Kimemia and S. B. Gershwin, \An algorithm for the
omputer
ontrol
of produ
tion in
exible manufa
turing systems," IIE Transa
tions,
vol. 15, pp. 353{362, De
ember 1983.
[12℄ R. Akella and P. R. Kumar, \Optimal
ontrol of produ
tion rate in a
failure prone manufa
turing system," IEEE Transa
tions on Automati
Control, vol. AC-31, pp. 116{126, February 1986.
[13℄ T. Biele
ki and P. R. Kumar, \Optimality of zero-inventory poli
ies
for unreliable manufa
turing systems," Operations Resear
h, vol. 26,
pp. 532{546, July{August 1988.
[14℄ A. Sharifnia, \Produ
tion
ontrol of a manufa
turing system with
multiple ma
hine states," IEEE Transa
tions on Automati
Control,
vol. 33, pp. 620{626, July 1988.
[15℄ problem with randomly
u
tuating demand," SIAM Journal on Control
and Optimization, vol. 25, pp. 1494{1502, November 1987.
[16℄ S. B. Gershwin, \A hierar
hi
al framework for dis
rete event s
hedul-
ing in manufa
turing systems," in IIASA Workshop on Dis
rete Event
Systems: Models and Appli
ations, August 3-7 1987. Sopron, Hungary.
[17℄ S. Gershwin, \Sto
hasti
s
heduling and set-ups in manufa
turing sys-
tems," in Pro
eedings of the Se
ond ORSA/TIMS Conferen
e on Flex-
ible Manufa
turing Systems: Operations Resear
h Models and Appli
a-
tions, (K. E. Ste
ke, ed.), pp. 431{442, 1986. Elsevier, Amsterdam.
[18℄ J. Perkins and P. R. Kumar, \Stable, distributed, real-time s
hedul-
ing of
exible manufa
turing/assembly/disassembly systems," IEEE
Transa
tions on Automati
Control, vol. AC-34, pp. 139{148, Febru-
ary 1989.