(i
_c, i
_w)
is a ZW reipe ar with the weight oft w
, that is, the maximal waitingtime possibleafter the ompletion of the intermediate.
(i
_c, i ′ ), (i ′ , i
_w)
are zero-weighted shedule ars that ensure thati ′
must startbe-tween the ompletion and wearing of the intermediate.
(i, i ′ )
is a UW reipe ar with weight oft pr
. This ar is needed so that the shedulear towards the task that follows
i
in the same unit ould start fromi ′
not onlyfrom
i
_c
.Figure5.7: Examplefor modeltransformation of LW stages
5.4 Comparison of approahes
The eieny of the algorithms from the previous setions is illustrated on a literature
example. Theexample features 5dierentsequential produtsand 6 unitsto besheduled.
The reipe graphfor the example is shown inFigure 5.8.
Figure5.8: Example for the omparison of LW/ZWapproahes
For the sake of the omparison, eah intermediate is assumed to have ZW storage
pol-iy. The investigated algorithms were ompared on 14 dierent ongurations with bath
numbers.
These algorithmsare:
sLP is the simple LP approah
aLP is the advaned LPapproah that opies the modelfor eah subproblem
aLP' is the advaned LP approah that uses onlyone modelasa global variable
Neg isthe approah with negativeweighted ars
Re isthe approahrelying on the reursive searh
Re+ is the reursive approahwith the extended positive ar additions
Eah test run were set within a 1000 s time limit. For the 11 smaller ongurations,
the approahes have not reahed this limitexept forouple of ases, and were able tond
the optimal solution. The data for the CPU times of the approahes are given in detail in
Setion C.2 , and illustrated inFigure 5.9
Figure5.9: CPUtime of ZW/LW approahes for the smallerases
The LP basedapproahes were usually 1or2magnitudesslowerthan the ombinatorial
approahes, among whihthe negative ar based proved tobethe most eient.
Forthe 3 larger ongurations all of the approahes have reahed the 1000 seond time
limit, and thus stopped. For these ases the omparison of the quality of the best found
(probably suboptimal) solution is important, and shown in Figure 5.10. The gure also
inludesthe2largestongurations fromthepreviousones, wherethe LPbasedapproahes
have reahed the limit.
Figure 5.10: Quality of reportedsolutions of ZW/LWapproahes forthe largerases
The aLP' approah ould not even report a feasible shedule for the larger ases. The
best solutions were reported by the negative ar based approah. The extended reursive
approah had very loseresults. The other approahes utuated.
In general, itan bestated that the most favorable approahseems to bethe one using
negative weighted ars.
Summary and onluding remarks
TheoriginalalgorithmsoftheS-graphframeworkwasdeveloped forUnlimitedWaitstorage
poliies. Inthishapter,dierentoptionsofextendingtheframeworkforLimited-and
Zero-Waitstoragepoliieshasbeenpresented andinvestigated. Theseoptionswere implemented
andomparedonawidesetofexamplesintermsofomputationalneeds. Theresultsshowed
thatanextended modelwithnegative-weightedars, andslightlymodiedalgorithmsisthe
most eient option.
Related publiation
•
Hegyhati, M., T. Holzinger, A. Szoldatis, F. Friedler, Combinatorial Approah to AddressBathShedulingProblemswithLimitedStorageTime,ChemialEngineeringTransations, 25,495-500 (2011)
Related onferene presentations
•
Hegyhati,M.,T.Holzinger,A.Szoldatis,F.Friedler,CombinatorialApproahto Ad-dress BathShedulingProblemswithLimitedStorageTime,presented at: PRES'11,Florene, Italy,May 8-11, 2011.
•
Hegyhati, M., T. Majozi, F. Friedler, Throughput maximization in a multipurpose bathplant, presented at: PRES 2008, Prague,Czeh Republi,August 24-28,2008Maximizing expeted prot in a
stohasti environment
InChapter 4thegeneralalgorithmfor throughputmaximizationwasintrodued. Formany
problems, however, several parameters of these types of problems are not deterministi.
Pistikopoulosetal.[104℄ gave alassiationofstohasti shedulingproblemsbased onthe
soureoftheunertaintyintheproess. Inahangingmarketenvironment,forexample,the
prie and marked demands an usually be onsidered stohasti. Both of these values an
have impat on the overall prot, as in ase of a overprodution, for example, the storage
of the surplus may result in additionalexpenses.
Obviously, in suh unertain irumstanes, it beomes ambiguous whih solution an
be onsideredasoptimal: the mostrobust one,orthe one withhighestexpetedprot,et.
Li and Ierapetritou[76℄ gave anextensivereview of the approahes that dealwith dierent
typesofunertaintiesinbathproesssheduling. Withoutattemptingtobeomprehensive,
the three main diretionof researh is illustrated inFigure 6.1.
Figure6.1: Classiationof approahes dealingwith unertainty
Preventive sheduling In this diretion of researh[72, 14℄, the shedule must be
given a-priori,and annotbemodied afterthe system starts,or the realization
of unertainevents. If thereisnoinformationabout theprobabilitydistribution
of theunertainparameters, areasonable objetive ouldbeaprodution
shed-ule, that exeeds a ertain prot, and have the highest level of robustness. On
the other hand, an estimation about the probability distribution gives room for
optimization towards the highest expeted prot.
Reative sheduling Reative approahes[92℄ assume an existing shedule, whih
gets disturbed by some unertain event. The objetive is to modify the rest
(unexeuted part) of the shedule ina way, that leads to the best available
per-formane. Unlike inthe previous ase, the optimizationisarried out, whenthe
system is already running, thus usually there is a very limited amount of time
to deliver the solution. As a result, heuristis are often favored for this
pur-pose. Onthe brightside, theapproah doesnot have todealwith any unertain
parameters, as they are already realized.
Two- and multistage approahes In a two-stage optimization approah[59℄, it is
assumed thattherearesomedeisionsthatmust bedonepriortothestartofthe
system, however, some deisionsan bealteredlater. Asanexample,a shedule
may has to be deided in advane, but the load of the units an be adjusted
after the unertain parameters our. In a multi-stageapproah,this onept is
generalized, and itis not assumed, that allof the unertain events realizeatthe
same time.
Naturally, there are a lot of developments, whih do not t into these three ategories,
some approahes, e.g., ombine preventive and reative sheduling. Sensitivity analysis in
itself is abroad area to researh,but itan alsopartiipate asa subroutine for preemptive
sheduling, espeially inase of heuristiapproahes.
ThefollowingsetionsintrodueanS-graphbasedapproahfortheshedulingof
through-putmaximizationproblemsonsideringunertainostanddemandparameterswithdisrete
probability distribution funtions. Setion 6.1 gives anexat denition of the problems to
be solved, and Setion 6.2 and 6.3 presents the S-graph approahes to address this set of
problems. In Setion6.4 anextensionis shown toontinuous probability distribution
fun-tions. Last, Setion 6.5illustratessome of the algorithmsvia an example, and draws some
onlusions.
6.1 Problem denition
The problemstobesolved are given by similarparameterstothoseof ageneralthroughput
maximization problem, i.e., eah produt is given with its reipe, along with the set of
equipmentunits andthe timehorizon. Atthis pointitisassumedthat thereisaone toone
relation between produts and reipes, i.e., there isno reipeproduing multiple produts,
and there are no two dierentreipesproduing the same produt.
There is, however, a set of unertain parameters for eah produt, whose probability
distributionis disretizedintojoint senarios. Thus, for eahdisretesenariothefollowing
parameters are given:
The objetive is to make deision about the number of bathes and provide a feasible
shedule in way to ahieve maximal expeted prot. Note that it is assumed that if the
atual produtionis higherthan the demand, the surplus isnot sold.
1
Basedonthepossibledeisionsrelatedtothesizesofeahbath,threedierentproblems
are identied:
Preventive problem with xed bath sizes whenthebathsizeforeahprodut
is given, and the only preventive deisionto be madeis todeide the number of
bathes for them.
Preventive problem with variable bath sizes is a more exible version of the
previousproblem,wherenot onlythenumberofbathes,but alsotheirsizesan
be altered inadvane before the unertain events realize.
Two stage problem where the bath numbers have to be deided in advane, but
the bath sizes an be alteredaordingly afterthe unertain events realized.
Setion 6.2 introdues the approahes that an solve these three dierent problems. In
Setion6.3,theseapproahesareextendedtotakleases,whenareipeanprodueseveral
produts, and the same produts an be produed by several reipes.
2
.
6.2 S-gaph based approahes
The three approahes to solve the problems presented inthe previous setion are all based
on the throughput maximization algorithm presented in Setion 4.1. The main algorithm
that examines dierent bath size ongurations and the feasibility tester an remain the
same, as inthe ase of the general throughput maximizationproblems. The subroutines to
seletaongurationandtoupdatetheset ofopenongurations anbealteredinorderto
1
Taking theopposite assumption would not alterthe struture of the algorithm, thesame approahes
wouldbeappliablewithmodiedinputparameters.
2
Theapproahforthepreventiveversionofthis asehasbeenpublishedin theliterature[72℄
ahieve higher eieny, however, any of the previously mentioned implementationswould
sue.
Thekeydierenebetween thedierentapproahesistherevenuefuntion,thatshould
provide the expeted prot, orthe highest expeted prot for a onguration, asit will be
disussed inthe followingsubsetions.
In order to simplify these desriptions, the following additional notations will be used
throughout this hapter next to the ones used previously: