3 Ro
3 Ro
3 Ro
Abstract—We consider a production control problem in a manu- of the resulting threshold control policy. In particular, they
facturing system with a failure-prone machine and a stochastic de- showed that the optimal production policy is characterized by
mand. The objective is to minimize a discounted inventory holding a critical number, known as the threshold level or the hedging
and backlog cost over an infinite planning horizon. The optimal
production control of continuous, stochastic manufacturing sys- point. The optimal policy requires the machine to produce at
tems with a failure-prone machine and a constant demand has been full capacity if the inventory level is lower than the critical
considered in Akella and Kumar [1]. However, the problem of op- number, to produce nothing if the the inventory level is higher
timal production control for discrete stochastic manufacturing sys- than the critical number, and to produce as much as the demand
tems with uncertain demands remains open. In this paper, we in- if the inventory level is equal to the critical number. Bielecki
vestigate a case where the exogenous demand forms a homoge-
neous Poisson flow. Primarily, we show that the optimal produc- and Kumar [3] dealt with a similar problem with a long-run
tion control for such a system is of the threshold control type. In average cost objective. It has been demonstrated as well in [3]
addition, the explicit form of production control policy and the ob- that the threshold production control policy remains optimal.
jective functions are provided. Numerical examples are included to Explicit analytical solutions of the stochastic failure-prone
demonstrate the results obtained in the paper and to compare with manufacturing system have so far been obtained only in Akella
the one in Akella and Kumar.
and Kumar [1] and Bielecki and Kumar [3]. With the assump-
Index Terms—Optimal control, Poisson demand, production tions of continuous failure-prone manufacturing systems and
planning and scheduling, stochastic manufacturing systems, constant demand, a number of researchers attempted to solve
threshold control policies.
more general problems than [1] and [3]. This line of research
has resulted in the existence and partial characterization of op-
I. INTRODUCTION timal production policies. For example, Van Ryzin et al. [16],
and Bai and Gershwin [2] provided an approximation of op-
B EGINNING with Kimemia and Gershwin [11], Akella
and Kumar [1], and Bielecki and Kumar [3], there has
been considerable interest in the problem of optimal production
timal feedback controls in the case of manufacturing systems
consisting of two or three machines in tandem. Hu and Xiang
planning and control of stochastic manufacturing systems. [10], Hu et al. [9], and Liberopoulos and Hu [13] extended the
Kimemia and Gershwin tackled a multimachine flowshop with Akella and Kumar [1] to the case of non-Markovian machine
unreliable machines without internal buffers. Kimemia and capacity process. For the long-run average cost objectives, Shar-
Gershwin demonstrated that a near-optimal control can be char- ifnia [19] extended Bielecki and Kumar [3] to the case of more
acterized by some threshold values (known as hedging point). than two machine states. Liberopoulos and Caramanis [12] fur-
Akella and Kumar dealt with a single-machine, single-part-type ther demonstrated that the Sharifnia’s approach applies when
system with linear holding and backlog cost. They assumed the machine transition rates depend on the production rates. In
that the machine capacity is a birth-death process, i.e., the addition, multiple product systems, especially systems require
time between successive machine failures and repair times either setup costs and/or set up times, were also under considera-
are modeled as exponentially distributed random variables. tion (see Gershwin [7], Gershwin et al. [8], Caramanis et al. [5],
They further assumed that the demand is a constant, and the Connolly et al. [6], and Sethi and Zhang [18]). For a complete
state and control variables are continuous. They obtained an survey in optimal control of stochastic manufacturing systems,
explicit solution to the problem and verified the optimality we refer the reader to the book by Sethi and Zhang [18] and a
recent survey paper by Sethi et al. [17]. Note that these results
are based on the relatively restrictive assumptions of constant
demand rate and continuous production systems.
Manuscript received January 13, 1999; revised October 25, 1999 and Feb-
ruary 4, 2000. Recommended by Associate Editor, Q. Zhang. This research is However, it is well documented that demand uncertainty is
partially supported by Grant RP3930033 of the National University of Singa- the one of major factors affecting the decision making in pro-
pore, and an Earmarked Grant from the Research Grant Council under Grant duction planning and control [14] and [20]. In addition, many
CUHK351/96E.
Y. Feng is with the Research Group, Enron Corp., Houston, TX 77002 USA manufacturing functions are discrete in nature. Hence, taking
(e-mail: youyifeng@enron.com). into account a discrete failure-prone stochastic manufacturing
H. Yan is with the Department of Systems Engineering and Engineering system with the uncertain demand will not only reveal theoret-
Management, The Chinese University of Hong Kong, Shatin, N.T., Hong Kong
(e-mail: yan@se.cuhk.edu.hk). ical properties and managerial implications but also open the av-
Publisher Item Identifier S 0018-9286(00)10662-2. enue of applications. Recently, Perkins and Srikant [15] studied
0018–9286/00$10.00 © 2000 IEEE
FENG AND YAN: OPTIMAL PRODUCTION CONTROL IN A DISCRETE MANUFACTURING SYSTEM 2281
continuous failure-prone manufacturing system with uncertain to ensure a stable system, i.e., the average capacity is no less
demand with a long-run average cost criterion. They demon- than the average demand
strated that hedging point policy is optimal with bursty demand.
In this paper, following the work of Akella and Kumar [1], (2)
we consider an optimal inventory/production control problem in
a stochastic manufacturing system where the machine capacity At the same time, we define the capacity and demand gap as
process is a birth–death process. In addition, we assume that the . By definition, .
demand is uncertain. Differing from Perkins and Srikant [15], We use to denote system states, where is
we consider a discrete manufacturing system with a discounted the inventory state, and is the machine state for
cost criterion in which the demand, the capacity of production, , i.e.,
and the processing time per unit are random variables. Specif-
ically, we show that the optimal threshold control policy can if the system is up
be specified by a critical inventory level. Moreover, we demon- if the system is down.
strate that this optimal threshold control policy minimizes costs
for all feasible production control policies over all the levels of Our aim is to minimize the operational cost on average dis-
initial inventory and machine states. counted over the infinite horizon, regardless of the initial inven-
The rest of paper is organized as follows: Section II intro- tory levels and the operational status of the underlying machine.
duces the system model and notations. An optimality condi- To this end, we can write the discounted cost over an infinite
tion is also provided at the end of the section. We evaluate horizon as
the threshold production control policies in Section III. In Sec-
tion IV, we first identify the optimal threshold production con-
trol policy and then demonstrate that this optimal threshold pro-
duction control policy remains optimal for all feasible control (3)
policies investigated. Section V provides numerical examples.
Final, we conclude the paper in Section VI. where is the level of inventory at under control
, and , respectively;
II. PRELIMINARIES is the discount factor.
The objective is to find an admissible capacity allocation, or
In this section, we provide background and notations. We will production control , which minimizes total inventory and
also characterize the optimality of the system with functional backlog costs over time. Note that an admissible control
equations which suffice to identify the optimal production con- requires that: i) is measurable in , ii) for
trol policies. , and iii) for . Specifically, the objective is
to minimize the discounted cost (3) for any initial state of the
A. Model Description and Notations
system , over , i.e.,
The underlying manufacturing system is subject to random
machine breakdowns. When the system is up, the machine can (4)
produce the product at a rate up to , where is the ma-
chine capacity. The time duration of making a unit of the product
is exponentially distributed. When the machine is down, the B. Optimality Conditions
system can produce nothing. When the machine recovers from The recent development in the point process intensity control
the breakdown, it will continue the incomplete work on the unit sheds new light on the problem of the production planning
left from the last breakdown. Because of the memoryless prop- and scheduling of stochastic manufacturing systems. In such
erty, the remaining processing time on the unit is stochastically systems, the production output has been considered as a flow of
equivalent to that of being processed from scratch. We further point process; the process is controllable by adjusting the pro-
assume that the durations of up and down periods are exponen- duction rate in the range of . In what follows, we will first
tially distributed with rates of and , respectively. The de- establish the specialization of the Hamilton–Jacobi–Bellman
mand process forms a homogeneous Poisson flow with a con- (HJB) equation for the intensity control of point processes
stant rate . The holding and backlog cost per unit of item over (see Brémaud [4]). The HJB equation is available when no
per unit of time are and , respectively. Therefore, we can interference is allowed during the middle of production. This
write the cost function as follows: point is interpreted by viewing the following relation satisfied
by an arbitrary feasible policy . For a feasible policy ,
if
(1) assuming that it takes rate at state
otherwise.
(6)
and when
The last equation is equivalent to
(7)
(9)
and
where is the determinant of matrix . The above determi- Hence, when , and , only is likely to be pre-
nant of matrix can be transformed into the following equa- sented. That is
tion:
(16)
(14)
and Substituting into the system simultaneously and com-
paring the coefficients, we get following equations:
(15)
if
(18)
if
These will be substituted explicitly into the expressions for the (22)
particular solution to (6) and (8), .
Analogously, when , we will be able to derive When , the obvious boundary conditions from (6)–(8)
a particular solution to (6) and (8) which we also denote as are
. As a matter of fact, it can be obtained from
the previous analysis by replacing with . Therefore, for
, we achieve a particular solution to the system of simul- (23)
taneous difference equations (6) and (7) as well. To sum up, we and
display the results of the foregoing analysis in the following for-
mulas:
(24)
if
(19) (25)
(28)
if
In view of (26) and (28), we further simplify (27). That is
if
(29)
(21)
FENG AND YAN: OPTIMAL PRODUCTION CONTROL IN A DISCRETE MANUFACTURING SYSTEM 2285
Thereafter, (26), (29), and (28) have comprised a system of si- D. The Values of Cost Function at the Levels Above
multaneous linear equations with regard to the three unknown When , (7) and (8) are valid when (6) is disabled.
variables . Solving it yields This makes the way of presenting the cost function in the levels
higher than different from that of those lower than . To ob-
tain the cost function for these levels, observe the characteristic
equation of the system of simultaneous difference equations (7)
and (8)
(30)
(34)
By factoring in the right-hand side of (34), we easily get the
solutions to (34) as
(31)
Thus, the general homogeneous solution to the system of (7) and
(8) is represented by
Substituting these expressions back into (7) and (8), we get the
following simultaneous equations:
Because of the boundary condition, when , both (21) and Note that determining for differs from and looks
(35) must coincide. That is easier than for . Considering (22) in and the
properties of , we have
(36)
and
(37)
Lemma 4.1: sions for it as well. This is the contribution of the following
lemma.
(40) Lemma 4.4: Let temporarily. Then
(41)
(43)
Proof: In view of (40)
Proof: See the Appendix.
In the following lemma, we present an equality which will be
useful in optimality analysis.
Lemma 4.5: For
Since
(42) (45)
B. Identification of the Optimal Threshold Control Policies where the function is defined as
We will show, at first, that the optimal threshold control policy
cannot be of a negative integer.
Lemma 4.7: The optimal threshold control policy cannot be
associated with a negative integer .
Proof: Recall the expressions for the cost function of a
threshold control policy with a negative number . In par-
Then is the minimum of .
ticular, for
Proof: This is straightforward from Lemma 4.9.
We are able to assert that exists finitely. This is actually
owing to a limit since .
However, the figure
if
if
For any Corollary 4.12: if and only if
(46)
This inequality implies that the policy associated with cannot Proof: In view of Corollary 4.6
be optimum since obviously the policy defined by per-
forms better than it, at least at all backlog levels.
As a result, we will be no longer interested in those threshold
control policies associated with negative critical numbers. Anal-
ogously, in their deterministic and continuous manufacturing Hence, if and only if . However, if
system, Akella and Kumar [1] pointed out the same result. and only if
To identify the optimal threshold control policy, we need the
following concept of unimodal function.
Definition 4.8: A function for a set of integers,
, is called unimodal if there exists a such that when This is equivalent to (46).
and is nonincreasing and when , On the basis of further analysis from Corollary 4.12, we
and is nondeceasing. would like to point out two possible situations where the
Lemma 4.9: is convex in . In other words, zero-threshold control is optimal.
is nondecreasing in . Corollary 4.13: When the capacity of production is large
Proof: See the Appendix. enough, irrespective of the backlog cost coefficient , the zero-
Lemma 4.10: is a unimodal function of . threshold control is optimal. In addition, when the backlog cost
Proof: It is easily checked out that coefficient is small enough, the zero-threshold control is op-
timal as well.
Proof: The last result is readily obtained from (46) since
where the last equality is due to (41) of Lemma 4.2. Hence, when
is large enough, (46) takes effect.
FENG AND YAN: OPTIMAL PRODUCTION CONTROL IN A DISCRETE MANUFACTURING SYSTEM 2289
In the next section, we shall show that the optimal threshold TABLE I
control is the optimum of all feasible control policies. SYSTEM PARAMETERS FOR TESTING EXAMPLES
TABLE II
COMPUTATIONAL RESULTS OF CASE 1, ( = 33:33)
TABLE III
COMPUTATIONAL RESULTS OF CASE 2, ( = 12:5)
Note that
and
Consequently, we get
We then substitute the last equality into the left-hand side of (47)
and get
It is readily seen that the last equation is just (44). Note that the
second-to-last equation has employed the following relation:
F. The Proof of Lemma 4.14 The last of the preceding derivations ensures that for
. Henceforth, the
For
Lemma follows.
The last inequality, which finishes this case, is due to the defi-
nition of . Now consider another case of . Note
2294 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 45, NO. 12, DECEMBER 2000
The foregoing expressions for can be employed in we will see it is an increasing function of . As a matter of
simplifying fact, for
where
It is induced that
where
Setting
FENG AND YAN: OPTIMAL PRODUCTION CONTROL IN A DISCRETE MANUFACTURING SYSTEM 2295
Utilizing Corollary 4.12, it is found that as shown in the equation In particular, . On the other hand, when
at the top of the page. Thus
[14] S. Nahmias, Production and Operations Analysis, 3rd ed. Chicago, IL: Houmin Yan received the B.S. and M.S. degrees
Irwin, 1997. from the Department of Automation, Tsinghua Uni-
[15] J. R. Perkins and R. Srikant, “Failure-prone production systems with versity, in 1982 and 1985, respectively, and the Ph.D.
uncertain demand,” Boston Univ., 1998. Working Paper. degree from the Faculty of Management, University
[16] G. J. Van Ryzin, X. C. Lou, and S. B. Gershwin, “Production control for of Toronto in 1993. He was an Engineer with CIM
a tandem two-machine system,” IIE Trans., vol. 25, pp. 5–20, 1993. Group, Semiconductor Manufacturing Operations,
[17] S. P. Sethi, H. Yan, H. Zhang, and Q. Zhang, “Asymptotic optimality Digital Equipment Corporation, Hudson, Mass. in
of hierarchical controls in stochastic manufacturing systems: A review,” 1991. He has been with the Department of Systems
Univ. Texas at Dallas, 1998. Working paper. Engineering and Engineering Management, the
[18] S. P. Sethi and Q. Zhang, Hierarchical decision making in stochastic Chinese University of Hong Kong as an Assis-
manufacturing systems. Cambridge, MA: Birkhäuser, 1994. tant/Associate Professor since Jan. 1994. His main
[19] A. Sharifnia, “Production control of a manufacturing system with research areas are operations management, stochastic models, simulations, and
multiple machine states,” IEEE Trans. Automat. Contr., vol. 33, pp. supply chain management. He has published in journals such as Operations
620–625, 1988. Research, Production and Operations Management, Discrete Dynamic Event
[20] E. A. Silver, D. F. Pyke, and R. Peterson, Inventory Management and Systems: Theory and Applications, Journal of Classification, Journal of
Production Planning and Scheduling, 3rd ed. New York: Wiley, 1998. Optimization: Theory and Applications, International Journal of Production
Economics, and IEEE TRANSACTIONS ON AUTOMATIC CONTROL.