Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Abstra t
We show that there are ranges of parameter values des
ribing an unreliable manufa
turing system for whi
h zero-inventory poli
ies are exa
tly
optimal even when there is un
ertainty in manufa
turing
apa
ity. This
result may be initially surprising sin
e it runs
ounter to the argument
that inventories are buers against un
ertainty and that therefore one
must strive to maintain a stri
tly positive inventory as long as there is
any un
ertainty. However, there is a deeper reason why this argument
does not hold, and why a zero-inventory poli
y
an be optimal even in the
presen
e of un
ertainty. This provable optimality reinfor
es the
ase for
zero-inventory poli
ies, whi
h is
urrently made on the separate grounds
that it enfor
es a healthy dis
ipline on the entire manufa
turing pro
ess.
In re
ent years, the goals of \zero-inventory" and \sto
kless produ
tion" have
attra
ted mu
h attention (Hall 1983). Su
h poli
ies enfor
e a stri
t dis
ipline
on the entire manufa
turing pro
ess that has several bene
ial
onsequen
es for
the overall produ
tion system.
In this paper, we examine the question of whether there are any
onditions
under whi
h a zero-inventory poli
y is a
tually optimal. In this
onne
tion it
should be noted that it is sometimes argued that a zero-inventory poli
y only
passes the
osts from, say, the assembly line to the parts manufa
turer, and
that, therefore, any advantages a
ruing from enfor
ing su
h a poli
y are only
side benets resulting from the greater dis
ipline governing the system, and not
dire
tly from the poli
y itself. After all, the reasoning goes, positive inventories
are used as a buer against un
ertainties in supply or demand, and so as long as
there is any un
ertainty whatsoever in the system, one should strive to maintain
(Please address all
orresponden
e to the se
ond author below). Main College of Planning
and Statisti
s, Warsaw, Poland.
y Department of Ele
tri
al and Computer Engineering and Coordinated S
ien
e Laboratory,
University of Illinois, 1101 W. Springeld Avenue, Urbana, Illinois 61801/USA.
a stri
tly positive buer. Thus, zero inventory levels
an only be optimal when
there is no un
ertainty at all, and that is never possible !
In
ontrast to this
on
lusion, we show in this paper that there are
onditions under whi
h a zero-inventory poli
y is a
tually provably optimal, even
when there is un
ertainty, and that furthermore su
h optimality of the zeroinventory poli
y results whenever the system is designed to be su
iently, but
not ne
essarily perfe
tly, e
ient.
The model with respe
t to whi
h we exhibit these
on
lusions is that of a
simple failure prone manufa
turing system produ
ing a
ommodity. The time
between failures is random and modeled as an exponentially distributed random
variable with mean 1=q1 , while the repair time is exponentially distributed with
mean 1=q0 . When fun
tioning, the system
an produ
e at any rate up to a
maximum of r units/time; while broken down it
annot produ
e at all. At all
times the
ommodity is being depleted at a demand rate of d units/time. The
total inventory
an be negative, whi
h just
orresponds to a ba
klog. Positive
inventories are assessed a
ost at a rate of
+ dollars per unit
ommodity per
unit time, while negative inventories are assessed a similar
ost of
. We seek
the optimal produ
tion poli
y for this system whi
h minimizes the long run
average expe
ted
ost in
urred per unit time,
ZT
1
[
+ x+ (t) +
x (t)dt
(1)
lim E
T !1
where x(t) = inventory level at time t (see below), x+ = max(0; x) is its positive
part, and x = max(0; x) its negative part. Let u(t) be the produ
tion rate
at time t, then
learly,
Zt
x(t) =
(u(s) d)ds
0
The quantity z , the optimal inventory level, is therefore key to the whole
problem, and it is expli
itly given by:
rq1 (
+ +
)
+ (r d)(q0 + q1 )
r dd
= +1 if
q1 q0
z = 0 if
((q0 =dd)
1 and r q d qd
1
(5)
rq (
+ +
)
1
log + 1
(q1 =(r d)))
(r d)(q0 + q1 )
otherwise.
(6)
(7)
Note that 1=q1 is the mean up-time and so r d=q1 is the maximum total
produ
tion in a mean up-time. Similarly 1=q0 is the mean down-time and so
d=q0 is the total depletion in a mean down-time. So if r d=q1 < d=q0 , then the
system does not have the
apa
ity to meet demand even if it produ
es at full
apa
ity when up. Thus, z = +1 under these
onditions.
So
onsider only the
ase r d=q1 > d=q0 . Then we have the interesting
situation that the optimal inventory level z is exa
tly 0 whenever
rq1 (
+ +
)
+ (r d)(q0 + q1 )
1:
This happens whenever q1 is redu
ed enough or q0 in
reased enough, i.e., whenever the system is made e
ient. Note that it is not ne
essary to have q1 = 0
or for q0 = +1 in order for a zero-inventory poli
y to be optimal. This is
somewhat surprising be
ause positive buers are usually maintained as a hedge
against future un
ertainty (about manufa
turing
apa
ity, in our
ase), and so
one feels that a stri
tly positive inventory should be maintained as long as there
is any un
ertainty, i.e., as long as q0 6= +1 or q1 6= 0, in the system. However,
there is a deeper phenomenon at work, as we explain in Se
tion I.
The above manufa
turing system is a spe
ial
ase of the multi-
ommodity,
multi-ma
hine, failure-prone
exible manufa
turing system
onsidered by Kimemia
and Gershwin (1983), whi
h in fa
t has been a key motivation for us. In Akella
and Kumar (1986), the authors have
onsidered a similar problem but with a
dis
ounted
ost
riterion,
Z1
E
e
t[
+ x+ (t) +
x (t)dt
(8)
0
where
> 0 is a dis
ount fa
tor. However, the average
ost
ase (1)
onsidered
here admits a very ni
e interpretation for the fundamental reason behind the
optimality of a zero-inventory poli
y; see Se
tion 1. The pre
ise proof of optimality, though, whi
h is ne
essary, and whi
h we provide in Se
tions 2 and 3, is
more intri
ate.
3
We hope that the expli
it expression for the optimal inventory level z , given
by (5-7), will nd use as a guideline. For example, the
ondition
+ (r d)(q0 + q1 )
rq1 (
+ +
)
I (t) = 1
I (t) = 1
I (t) = 1
(9)
(10)
(11)
(12)
The only quantity to be
hosen is z , and we shall optimize over its possible
values. (This is a
onsiderable restri
tion of the
lass of all poli
ies, and in fa
t
the proof of Se
tions II and III is needed pre
isely be
ause one does not know
a priori that other types of poli
ies
an be ruled out.)
We will assume that the
ombined pro
ess (x(t); I (t)) has a steady state
probability distribution,
P z (A; i) := tlim
!1 P rob(x(t) 2 A; I (t) = i)
when the poli
y (9-12) is used. Dene
The average inventory
ost (1)
orresponding to this poli
y, whi
h we denote
by J (z ),
an then be
omputed as the expe
ted
ost with respe
t to the steady
state distribution, i.e., it
an be written as,
Z0
Z1
z
J (z ) =
jxjQ (dx) +
+ xQz (dx):
(13)
(14a)
(14b)
Note from (11 and 12) that if the inventory level xz (0) starts with a value larger
than z , then it is depleted at rate d until it hits z . Thereafter (10 and 12) ensure
that it never again rises above z ; see Figure 1. Hen
e, xz (t) z for all t large
(15)
A + :=
fx : (x ) 2 Ag:
To see (15), suppose that fi g are the swit hing times of I (t), i.e.,
If fxz (t); I (t)g and fxz+
(t); I (t)g are the pro
esses
orresponding to the
hoi
es
of z and z +
, with initial
hoi
e of
J (z ) J (0) = jz j for z 0:
So,
1; z ))
J (z ) J (0) = [
+ z
zQ0 ((
+
( + (x + z ) + x)Q0 (dx);
(16)
forz 0:
The derivative at 0 is
d
(J (z ) J (0))z=0 =
+
Q0 ((
dz
1; 0)):
Q (( 1; 0));
(17)
i.e., the
ost of the probability mass at 0 is large enough. Moreover, for z > 0,
(d=dz )[J (z ) J (0)
an be
omputed to be,
d
[J (z ) J (0)z0
dz
+
=
(1 )
+ (
+ +
)Q0 (( z; 0));
whi
h is also positive when (17) holds. Hen
e (17) is both ne
essary and su
ient for z = 0 to yield a minimum of J (z ).
The basi
reason for the optimality of an exa
tly zero inventory poli
y should
now be intuitively
lear.
Using Q0 (( 1; 0)) = 1 , we
an rewrite (17) as
(
+ +
)
:
(18)
Next, we need to
ompute . This is done by determining pzi () and in
steady state. (Re
all that pzi () is the density fun
tion of P z (dx; i) in ( 1; z )).
t
Let pz;t
i () and denote the
orresponding quantities at time t. Considering
the probabilities at time (t + h), for small h > 0, we have:
Z
z;t+h
p0
z;t+h
p1
(x)dx = (1
q0 h)
+q1 h
Z
(x)dx = q0 h
A+hd
(r
= (1
q1 h)
q1
7
d)h
p0 (y)dy
for z 62 A
t+h
pz;t
1 (y )dy + o(h)
z;t
A+hd
+(1
pz;t
0 (y )dy
h)t
(r
Z
z
d)h
pz;t
1 (y )dy + o(h);
z
(r
d)h
pz;t
1 (y )dy + o(h):
+h
t+h = t . Taking the limit as
In equilibrium we should have pz;t
pz;t
i
i and
h ! 0 in equilibrium, and pro
eeding formally through,
Z
Z
z
pz0 (y + hd)dy
p0 (y)dy =
A
A+hd
Z
dpz0 (y)
z
=
p0 (y) +
hd dy + o(h);
dy
dpz0 (x)
=
q0 pz0 (x) + q1 pz1 (x)
dx
dpz (x)
(r d) 1
= q0 pz0 (x) q1 pz1 (x)
dx
(r d) z
p (z ):
=
q1 1
( d)
q1
pz0 (x)dx =
q0 + q1
(19)
(20)
(21)
(22)
q1 (x z)
e
for x z
q0 + q1
q1 d
pz1 (x) =
e(x z) forx z
(q0 + q1 )(r d)
d
;
=
q0 + q1
pz0 (x) =
(23)
(24)
(25)
:=
q0
d
q1
:
r d
(26)
> 0;
(27)
+ (r d)(q0 + q1 )
rq1 (
+ +
)
8
in (5). When this
ondition is not satised, straightforward evaluation of the
minimum of (16) using (25)-(27) yields the minimizing value of z presented in
(7).
We summarize below the
omputed values for the optimal
ost J (z ):
q1 rd
when z = 0;
q0 d q1 d)
!!
,
+ d
+
+
=
+
q0
q1
q0 + q1
d
r d
+
log
+ (rqr 1 (
d)(+q
+)q ) when z > 0:
0
1
J (z ) =
(q0 + q1 )(q0 r
dx(t)
= (x(t)); I (t)) d
dt
(28)
where I (t) is the Markov
hain denoting whether the system is up or down.
Some
onditions are needed on the fun
tion for there to exist a solution
of the dierential equation when (; i) is dis
ontinuous (as the optimal poli
y
is). We will assume the following
(29a)
(; 1) is measurable, and for every (; ) 2 R2 with 0, there exists a unique
fun
tion y (t; ; ) whi
h is absolutely
ontinuous in t,
ontinuous in (t; ; ),
and whi
h satises
y (t; ; ) = +
lim
t#
(t
)
Z
( (y (s; ; ); 1)
d)ds
for t ;
(29b)
(29 )
lim
T !1
E [x2 (T ) = 0:
(29d)
We shall say that a poli
y is admissible if it satises (29a, 29b, 29
and stable if
it satises (29d).
The reason for imposing (29a) is
lear. The
ondition (29b) guarantees
a solution of the dierential equation (28) while still allowing dis
ontinuous
fun
tions ; see Akella and Kumar where su
ient
onditions are given. The
ondition (29
) rules out some pathologi
al fun
tions and allows the denition
9
dW (x; i)
d
dx
qi [W (x; i) W (x; 1 i) +
(x) J = 0
for i = 0; 1 and all x
dW (x; 1)
dW (x; 1)
= min [u d
[ (x; 1) d
0ur
dx
dx
2
jW (x; i)j k1 x + k2 for some k1 ; k2 :
[ (x; i)
Then
lim
T !1
E
1
(x(t))dt = J lim 1T !1 E
T
Z
0
(30a)
(30b)
(30
)
(x(t))dt
Proof:
= qi h
f x + (y; i)dy
0
0
Zh
ds
+
(y; 1 i)dy)
h
s
Zh
+(1 qi h)f (x +
(y; i)dy) + o(h);
where :=
10
"
df (; i)
= qi f (x; 1 i)h + (1 qi h) f (x; i) +
d
+o(h) for somex x + h:
(yi ; i)dy
lim
[ (x(t); I (t)) d
dx
+qI (t) [f (x(t); 1 I (t)) f (x(t); I (t)) dt
= E [f (x(T ); I (T ) f (x(0); I (0)):
E
Now we apply the above formula to W noting that sin
e jx(t)j is bounded by
for 0 t T we
an modify W outside this range so that it
vanishes outside some
ompa
t set. So we have
"Z
T
dW (x(t); I (t))
E
[ (x(t); I (t)) d
jx(0)j + (r + d)T
dx
E [W (x(T ); I (T )):
Dividing by T , taking the limit and using (30 ) and (29d) to see that
E [W (x(T ); I (T ))
T
yields
lim 1T !1
E
RT
0
! 0;
(x(t))dt
T
J :
+ (r d)(q0 + q1 ) q1 r(
+ +
) 0
J =
q1 rd
(q + q )(q r qo d q1 d)
0 1 0
W (x; 1)
W (x) = W
(x; 0)
(q0 + q1 )
x2 11
=
2(q0 r q0 d q1 d)
r
x 0
(q0 r q0 d q1 d) 1
rd
0
for x 0
+
(q0 + q1 )(q0 r q0 d q1 d) 1
(q +q )x
rd(1 e 0d 1 )
+ 2 1
q1
x 1 +
=
q0
2d
(q0 r q0 d q1 d)(q0 + q1 )2
q1 r
x 1
(q0 + q1 )(q0 r q0 d q1 d) 1
rd
0
+
for x 0
(q0 + q1 )(q0 r q0 d q1 d) 1
(x; i) = r if x0; i = 1
= d if x = 0; i = 1
= 0 otherwise.
T !1
E (x2 (T )) = 0:
(31)
This
ondition essentially states that the inventory grows more slowly than the
square-root of time, and it is
lear that any reasonable poli
y that we wish
to implement will have to possess this property. Here, we need to prove this
stability property in order to invoke the su
ient
onditions for optimality given
in the Theorem of the pre
eding se
tion.
Just for
larity, we fo
us on the z = 0
ase; the details are similar in the
ase of z > 0.
Let fn g be the sequen
e of stopping times at whi
h the pro
ess (x(t); I (t))
hits (0; 0), that is
F (s) := P rob(n+1 n s)
N (t) := supfn : n tg
be the number of renewals whi
h have taken pla
e prior to time t. The quantity
(t) := t N (t)
is the
urrent life whi
h denotes the elapsed time sin
e the last renewal. Sin
e
x(N (t) ) = 0 and the inventory
annot deplete at a rate faster than d, we have
0 x(t) (t)d
and so
x2 (t)
(t)d :
2
T !1
E (2 (T )) = 0:
(32)
We therefore need to examine the se
ond moment of the
urrent life random
variable (T ).
13
Fix l and
onsider a typi
al renewal period, whi
h is the time interval
[l ; l+1 . Let
(33)
(34)
Ym := 2m
2m
and Zm := 2m
2m
(35)
be, respe
tively, the lengths of times whi
h I spends in states 0 and 1 before
a swit
h o
urs. Note that fYm g is a sequen
e of independent and identi
ally
exponentially distributed random variables with mean 1=q0 , while the fZm g are
independent and exponentially distributed with mean 1=q1 . When I = 0 the
inventory depletes at rate d; when I = 1 the inventory builds up at rate (r d)
(until the inventory returns to 0, whereupon it stays at 0). So
!
m
X
(r d)Zi dYi for 2m < l :
x(2m ) =
i=1
Hen e if
m
X
2m := min 2m : [(r
i=1
then
d)Zi
dYi 0 ;
)
l+1 l
m
X
= min 2m : [(r
i=1
Sin e
r d d
0;
q1
q0
the Law of Large Numbers shows that l+1 < +1 a.s. Hen
e F (s) is a proper
E [(r d)Yi
dZi =
distribution.
In what follows, we will assume without loss of generality that 0 = 0. Let
GT (s) := P rob((T ) s)
be the probability distribution of the
urrent life at time T . From renewal theory
(see Karlin and Tyalor, p. 193, 1975) we know that
ZT s
GT (s) = F (T )
[1 F (T y )dm(y )sT
= 1
s T;
14
where
m(t) := E [N (t)
is the mean number of renewals. Hen
e we may
ompute the se
ond moment of
(T ) as follows,
E [2 (T )
Z1
=
= T [1
2
s2 dGT (s)
F (T ) +
= T [1
F (T ) +
= T 2 [1
F (T ) +
= T 2 [1
F (T ) +
Z
Z
= T 2 [1 F (T )
ZT ZT y
+
2y
[1
0
s2 dGT (s)
2ydydGT (s)
ZT
2y
dGT (s)dy
0
2y [GT (T )
= T 2 [1 F (T )
ZT
+
(T z )2 [1
0
GT (y)dy
F (T
z )dm(z )dy
F (T
z )dm(z ):
(36)
We now intend to apply the Key Renewal Theorem (see Feller 1968) to obtain
the limit of (36) as T ! 1. To do so, we need to show that
F is not latti
e.
lim T [1 F (T ) = 0:
T !1
The distribution fun
tion F has nite mean.
The fun
tion t2 [1 F (t) is dire
tly Riemannintegrable:
2
(37a)
(37b)
(37
)
(37d)
In the lemma whi
h follows we will prove all four properties. Then, upon
applying the Key Renewal Theorem to (36), we will have
Z
1 1 2
lim E [ 2 (T ) =
t [1 F (t)dt < +1:
T !1
So (32), and the stability property (31) will follow readily. To
omplete this
argument, we thus need the following lemma.
Lemma. The properties (37a-37d) hold.
15
Proof:
F (t) is ontinuous:
(38a)
F (t)
t k
3+
(38b)
Clearly (37a) follows from (38a); (37b) follows from (38b); and (37
)
R follows from
(38b) through the fa
t that the mean of F
an be
omputed as 01 [1 F (t)dt.
Finally, (37a) shows that t2 [1 F (t) is Riemann integrable over every nite
interval, and (38b) is then a su
ient
ondition for dire
t Riemann integrability
(see Feller).
To prove both (38a) and (38b), let l := 0 and dene fm ; Ym ; Zm g as in
(35). Let x_ (t) = d for 2m t2m+1 and x_ (t) = r d for 2m 1 t2m ,
whi
h are, respe
tively, the intervals in whi
h I (t) = 0 and I (t) = 1.
For a 0, dene
be the probability distribution of the stopping time when x rst rises to a level
a. Clearly,
a
;
r d
sin
e x
annot rise more rapidly than at rate (r d). By
onditioning on 1
and 2 we get the renewal equation
F (t; a) = 0 for t
F (t; a) =
+
rt dt a=r
Z1
q0 e
a+1 d=r d
Z rt dt a=r
0
q0 1 d
q1 e
q1 (2 1 ) d(
q0 e
q0 1 d
(a+1 d)=(r
1 )
d)
qe q
F (t ; a + d + r r)d(
1
1( 2
1)
1 ):
(39)
Let the right hand side above be dened as (HF )(t; a), i.e., the a
tion of an
operator H a
ting on fun
tions F (x; y ) with x t. Clearly,
sup j(Hf )(x; w)
[1 e
q0 t sup
x t
w
xtw
and is therefore a
ontra
tion with respe
t to the sup norm. Moreover, H preserves
ontinuity and so the unique solution F of (39) is
ontinuous. Noting that
F (t) is the
onvolution of F (t; 0), with a mean 1=q1 exponential distribution,
we have proved (38a).
To prove (38b), note that by the Prin
iple of Large Deviations (Varadhan
1982) the quantity
n
1 X
1
Y
1 P
n i=1 i q0
n
!
1 X
1
Z
and
n i=1 i q1
de
reases exponentially in n. This implies that
n
X
1
1
+ ++
(Yi + Zi ) n
1 P
q0
i=1
q1
n
X
q1
q0
r d d
> ;
q1
q0
there exist , > 0 so that
1
(r d)
q1
q0
+ > 0:
q0
q1
4 CONCLUDING REMARKS
We have shown that zero-inventory poli
ies
an be optimal even when there is
un
ertainty in manufa
turing systems. This adds support to the use of su
h
poli
ies, whi
h are
urrently the fo
us of mu
h attention due to the greater
dis
ipline that they enfor
e on the manufa
turing system.
The parti
ular problem for whi
h we have obtained the solution is somewhat
restri
tive. It would be useful to relax several of the features. For example, the
demand rate, whi
h is
onsidered a
onstant here, may be time-varying and
featuring linear growth as well as seasonal and
y
li
al
omponents. It is also
important to take into a
ount the
hange-over
ost when swit
hing produ
tion
rates. Also, it would be useful to
onsider more general bat
h manufa
turing
problems.1
In a dierent dire
tion, it is of mu
h interest to examine the
omplex model
of a multi-ma
hine, multi-produ
t
exible manufa
turing system featured in
Kimemia and Gershwin (1983) to show that there are ranges of mean times between failures and mean repair times under whi
h zero-inventory poli
ies remain
optimal.
ACKNOWLEDGEMENTS
The authors are grateful to the editor, asso
iate editor and the two referees for a
number of
onstru
tive suggestions in
luding those in Se
tion ??. The resear
h
of the se
ond author has been supported by the U.S. Army Resear
h O
e
under Contra
t No. DAAG-29-85-K0094, and by the Joint Servi
es Ele
troni
s
Program under Contra
t No. NOOO14-85-C0149. The work of the rst author
was supported by a Fulbright S
holarship while he was visiting the University
of Illinois.
REFERENCES
Akella, R., and P. R. Kumar. 1986. Optimal Control of Produ
tion Rate
in a Failure Prone Manufa
turing System. IEEE Trans. on Automati
Control, AC-31, 2, 116-126.
Feller, W. 1968. An Introdu
tion to Probability
Vol. II, John Wiley & Sons, New York.
Hall, R. W. 1983.
1 The
18
19
J =
W (x) =
=
=
+ d
0
for x z
q1 (q0 + q1 ) 1
where
A1 =
q1
q1
r d r d
q0 q0
d
d
; A2 =
rq1 (
+ +
)
+
+ d
+ q0
q1 log[ +
q0 + q 1 ( d r d )
(r d)(q0 + q1 )
W (x; 1)
W (x; 0)
1 Am 1 xm !
1 An 1 xn !
1 An 1 ( z )n ! X
X
+ ( A1 z) X
+
1
1
1
b1
e
n
!
m
!
q
n!
1
m=1
n=1
n=1
!
!
+
1 An 2 xn
1 An 2 ( z )n
X
X
+
0
1
1
+
b1 +
b1
A1 1 +
n
!
n
!
q
q
1
n=2
n=2
!
1 An 2 (x z )n
X
+
1
+ b1
A1 01
n!
q1
n=2
+
+
(x z )
d
0
for 0 z
q1
q1 (q0 + q1 ) 1
1 An 1 (x z )n !
1 An 2 (x z )n !
X
+ X
0
2
2
+
b2
1
n
!
q
n
!
1
n=1
n=2
q1
qd0
d
q1
d
q0
d
; b1 =
r d
1
; b2 =
and
z =
( qd0
rq1 (
+ +
)
:
q1
+ (r d)(q0 + q1 )
r d)
(x; i) = rifxz ; i = 1
= difx = z ; i = 1
= 0 otherwise:
log[
20
1
d
1
d