Vous êtes sur la page 1sur 124
SEQUENCING AND SCHEDULING Copel oak Ne Melee iadio i oie) ash Cua) Table of Contents Preface (Cupiee 1 SCHEDULING PROBLEMS. seston of al scheduling (Cuprer2 OPTIMALITY OF SCHEDULES BI" ateder Regular measuies and emt sheds 24 Referenceand fares eading (Chapter 3 SINGLE MACHINE PROCESSING: BASIC RESULTS 32 Pamutton ches 33 Shona Prong Tne whedlg 35 Moot tlgtin “ Table of Contents Tle of oats “ ‘Caper 4 SINGLE MACHINE PROCESSING: PRECEDENCE Caper HARD PROBLEMS AND NP-COMPLETENESS CONSTRAINTS AND EFFICIENCY Or tarot a7 4.1 ntact rm $2 Whtina ps wy song pate i 42 Rogues ig fs st 83 Melancr? md is 43° Lanter salgotin. 3 84 TheNP-conpltenen figura ins 150 44 SPTveurncin mje woud cts. 0. 38 8S Retoencer sn ese Is 45° Pdpschedllict wih pecan ind #0. 38 95 Probie is $6 Releases end he sai. o 43° Poti o ‘ape 0 UEURISTICMETHONS. GENERAL APPROACHES To troduction 1s (Gapters CONSTRUCTIVE ALGORITHMS FOR FLOW-SHOPS AND 102 Sead peronchiges 136 sosstiors 103 Nedied or sponte meena pg tide. os sa Intodutin “ 104 Negiehoo mach eciae re 32, Somme oneal ‘ 103 Plowsop sedan haat 1% 53 Ieasoas soni forthe Fe poten © 106 eens and ure ding in 54 ate’ grt thn G]Poe elon % 107. Robins mm 53 Abpelemet ten) Ran ik 3 56 Ate’ pupal sano Bey PiPee pois 7 ‘Caper 11 HEURISTIC METHODS: WoRST CASE BOUNDS 53 Refereed father eaing. n Ti taredcien 1% Sh Polen 2 112 Stedlng depen obo ial mies me 113. Peon punsinstd N-cerpleoas te DYNAMIC PROGRASDIING APPROACHES 1A Woot coe malas expected peormance 19 oe © M13 Retenor ad terrng rt Th proach Hd Karp = 116 Problems ‘1 3 Gamputtinal ups of rani roping 2 4 Dynan poganmg subject 9 presece anacans 98 Gaps 12 A BRIEF SURVEY OF OTHER SCHEDULING PROBLEMS 5 Reforms and utr eding ia 21 invasion » $8 Robie i 122 recede const td suc coms in 123 Varacs ofthe sat eb-bop pole is ene 00s 124 Dyan nd tcc heady me eter iumment cee ie 125 Awembly ine balancing: pot management ad ie 12 Dynan propamning dit inate ues 10% tag ecures 206 13 Alloway camp Fr z 4 Sue perp ht th banged bach appa. 119 Sle probes x 1S Reward rer aig bs 36 Petre so 20 Stal aes 2 (aper 8 IWTEGER PROGRAMMING FORMULATIONS Sr neaeton 0A 82, Wapernep prop ies fnPiga pa 132 83 Referener a eer ada as Sate de, 18 BA Prtiene 135 x nn ‘inet Ac ally see Beco leave 1 work wih my publi Fee amay an a the ffisency ad tact une eve (pen amare th ppputin of Cs book SINON FRENCH aint of Manchester ee Mar 1881 7 Chapter Scheduling Problems eee 1a INTRODUCTORY EXAMPLE Alay, Bertie, Chae and Dighy shire a at, Fach Saturday they have four teespaper delvered” The, Fanci! Times, the Guardian, the Daily [apr and the Sun, Bei» snall-nindedceatore of babi each treet ofthe ft its on reading al the paper ia hs ova particular Suter Aly kes to begin withthe nancial Tames for hour. Ten he {temo he Cuando aking 30 shnutes glances ot he Dal Expres Tor 2 Inintes and finishes with 8 minutes speat onthe Sun. Bere prefers tO Regimen he Guardian taking 1 ou migutes hen he reads the Day [spe for inate, the Financial Time for 28 mutes and the Su for {ites Chaces begins by radg the Daily Expres for S minutes and fotloms ths with the Guarda for 13 minutes, the Pnancial Temes for 10 vite andthe Su for 30 rites. Finally. Dighy warts with the Si {hing T how 30 nutes, beoee spending | minute eah onthe Financial Times. the Guidian, and the Dal Express i. tha order. Each is 30 insistent spon hi partir reading order that he wil wat for his next paper tobe fee rather than select another, Moreover, pone wl release Faper nthe bas finshed Gren that Algy ges wp a 8.30 am, Bertie En Charest SS any and Digby 0930 9a tha hes can manage ovat shoe, save, and eat beakas while eadig 3 newspaper. and {wen ot each nuts on reading alte newspapers before going out what ‘Stheearet time that the four them can eave together for day in the snes” The problem faced by Algy and his tends. namely in what oder sould they rotate the papers Between teasetes so that al the reading s ned s soon as poste, yp ofthe scheduling problems that we Stall be consdcting Before describing te genet sux te of hee rob tems and ging some exunples tha re, perhaps more FCRvant £9 OF modern society ts worth examining hi ena fart Tet x ite the te tn more compact for, asin Table 11. How night we acl is pblen? Peep wil be eases 1 beginby exp Scheduling Problems ent Table 1.11 dats forthe newspaper rang prem or eer Gestp A Bitlng Order Ant Te ny a0 ra 0B s (60) GO) (2) oy ene 84s a oe fm & c) @ a8 to cinies 44s fe 2 fs @ a) aH by Pay 980 @ ge 6) be @» oO Oo © ing what scant by a reading sehedle. Quit imply it 2 prescription of the onder ahi the papers rotate between readers. For instance, one fossle schedule i shown in Table 12, where A, B.C. D denote Aly Bente, Charles, ad Dighy respective” Thus Algy has the Financial Times st bere it pasits to. Digy, then to Charles, and finaly 0 Berte, Silay the Guardian paces Ptwwen Beri, Charles, Aly. and Digby in that order. Aad so on “Table 2A posibe reading schedule Read by Peper dda mr A DOC OB G. - ¢ « OD ba Oo OOD s ow 88 We may werk ou ow long this eadiag schedule wil take by plotting « simple dageam called « Gant Diagram (Se Fg 1.)- this we pl four {ineaves One foreach newspaper Bkcks are paced ove the akes to inca when an by whom particule pupers ate ead Fr insta. the biock in the top lef hand corer ince that Alay reads the Financial Times from 830 10930 draw this dingram we have otated the papers pha “te [ie els] ae ‘nthe order given by Table 1.2 withthe rests hat eae reer flows ‘bs dosrd reading Order. This resem mete that for some the me oper ae lett uncead even when ther ae people who ar free amd have Pot read them yt, they mu emaim need antl someone eras) 4 08 {hom next. For tates, Berto could have the Paola! Tene at {00 41m but me wane te Daly Expreirt nd lees the Pinca "mer Simary the schedule i ano esponsble for idle time of the ‘ers Betucen 10,15 and 11.02 Chale wats forthe Fnontal Tr, Ueick foal bt the ast mina sot Being rea, but hares cant ate ‘he paper unt after Dighy Becatne of the sched teu. om the Gate dgram. eh ealest that all our can go out together 11 Stam. fey ase hecho. So the next question Fang snivcan we ind them a better sched, eone tt alls them t 0 fare Tram leaving that question for you to consider inthe fist se of Probleme Homevcr: before attempeng hose we should sna fasble Sa infeasile sees In Table 12 I snply give you 2 schedule and we saw inthe Cait gram that ths sched wold works tts posable for Aly an his Mates topes the papors amongst thermelses in hs order Bu uppnse "had given You the Schedule shown in Table 13 litle thous shows hats shes wil not work Aly fs offered the Sun fir, but he does ot want it ul he has Yea he other tree papers He cannot have ny of ‘ Scheduling Prolene ent “Table {3A infeasible shale these uni igh a finshed with them and Dighy wil ot start hem wl ea sun: which be sanaot have unl gy Bas net me ng Icy the reading orders gven i Table 1 ate called eed canraints, Any sedate which x compatible with these Uo fe Thue Table 12 ges a feasle schodul. Ufesble ced Fema in Table 13, ae incompatible withthe hi ed ee tinal. toe aceetabie a solution 10 a seul problem mate fel 12 PROBLEMS reset these probieans now before rang any further. eis ue that aero ie or no gakdanes on how they ae 1 e solved, have aa a ent eason Seeding «subjects wich the problems re pot ea. They ae on the contrary, among the ardent soo ou wil not apresate this without trying some for youre eitep hem wrth onnportant sal soe thes for yoo sort See tats important hat yo should discover thei difcoy 7%, We following sehedle eae for Aly ad bis tind? Sec.13) ob-Shop Scheduling Problem 5 2. How muny diferent schedules, fase or intense, ar there? 5 hats thecal ime tha Aly an his fend can leave forthe 1 Diaby decides that the dels of «day in she country at wot for th Intead he wl spene the morning ed. Oly when the others hare re got up aod reod the papers What the caries te that AB). Brie, and Chaees may fave! S Wetner of not you have solved Problems 3 and 4 consider Row so would recone the etfs possible departure tine, Need you com Pak Teeapicny mth tose all the other lease sehedalss, of can you about th proves of complete enumeration ofa the posses? 1.3 THE GENERAL JOB.SHOP SCHEDULING PROBLEM If the eheory of schedaling were simply concerned wih the efficient Fading of newspapers then of sours, ao one would study began with Pee olen that you ight mest and attempt 1 solve a scheduling (ble unhindered by the defetions and notations tht ae usally ren, tine her Zome to induce these definions and that 10% Ton The tesmologs of steduling theory asose inthe processing 2d ‘Minuicrenng.mdustdes. ‘Thar we shall be talking out jobs and Thathinen eventhough in some cases the objects releted to be title vrinten to ees jr oc machines. For stance. in the example of he ast {Elton shall sce that Ale, Bere, Charles and Digbvare jobs, whist Fe IABie te machines However, thts atcpaing We Begin by {ining tne genera joes problem. Soriy we shall bow tha is stuc cree say scheduling problems ang busines, compan, £0 ‘mene and the socal services a Wel as those i ny We chal suppor that we have Jobe ye.) 10 e processed theough mr machines (MM. -Me} Some authors particulary those Mc doinputerschedling e884 machines processors. We shall TERRES That cach job mat pom theough each machine once and one oa Tae processing of ¢ job on a wachne Is elle an operation, We shall evote te optaton on heh ob by ehe th machine bye, echo ‘al constrains demand tat ens job shoul be prosessed through the ‘fashierin a parlor ores, For general job-shop prblens here fe 0 “ation upon the form the feaologal constraints. Each jb haste aaaassseondce and thsmay bear no elation to he processing orser ‘iu her job, However, an moran special ase aes when al the fase the wie prosesing ordre. ln soc cumstances wes hat We Eine» fow-sho problem (heeaue the jobs flow between the machines Thea onde) Phs diction between jabalops and ow-shops souk! he made lear by the examples below 6 Schedating Problems {ent ach operation ote cern eng of time, the processing me 2 pera We denotes by py convention we fhe i any rete fea or sete the echt to proces ths job, We aso eee iy ane regu wo transport the jo tothe machi, We shall sacs ae tp ne fixed and How advance. Tis Brings wf a aa pLcestacon, which me stall make houghout hs Book. We shall rperiatpar eve manerie quantity deterministic an Kpown (0 the shedale. Se sh alo au thatthe machines ae always valle, but me stall tot nescaniy sume the sare for jobs, Some jobs ay not become et ut ater the seneling period Has started. Weal denote by TOA or reese date of he th job, Lethe time at which J fsmes nate for processing a ara prolere io find a sequence in which the Js pass etmoce the machines, Wik (a) compatible with the texhnologeal conan, Le. & feasible ‘hecule and (0) Spnimal wh respect some rteron of pernmance We shall discus possible rtri of pestormance showy: fst we stall oom smi purl examples, Indastilexaaples et ctrng em not engaged in mans pton of sina ies a ae endl problems at essa a that ofthe jb onc Tency wi have he own rout tough He varus work areas and pres athe factory. tn the lth indy diferent tes bane i a mets ettg sewing. prenang end Decking ta tec ile Cr et eres pind poses thou the set of rollers nts ON er order and wah ae cwn prior temperature and prec Pap nhc pray ins the tne pent it tpeseting a book wil Tere chtnlength numberof horations, and 9 om the He sent sere et pte wit Jepend on bth eth and ne number printed ae aera on binding wil depend nthe number pase and Ls the a se nackaging mil epend bork onthe numer pined a the dir rea fa anetcr wt mos sched the production of various Tek Apt hic peeling. printing. binding ond packaging depart aoa eee ours fowshop pele: for eath department 9 & ae eal hc ps Le the hooks, Mow from typesetng to pring Singing to packaging. Se Peewee in schedaling will vary fo fi 0 fn and often rom day dan espe bea wed Be to mina am eg eel of eis synth pense and machines were seo 6 see 14] a eee e perhaps t would be o achieve certain contracts! get dates simpy 0 Fann the work ap soon as posible Tete and some other abjetves will FeShewoe in det section 1S ad al of Chapter 2 Ay a ends Ae four ob four machine problem The jobs—Algy ert. Chae. TRI Dighy crust be schedules tea te four mahiresthe Fnac Sethe Guandin. the Daly Expres. sn the Suni ont 0 min tects nctntie athe pocai ang ed ‘Thee ave ceady tins, ely. the tates at wb ATgy et up. Nok Thats this cumple the technolopeal contains dot derand the Ma (Paley read the paper in the sie onder ecu ofthis we have Jo ‘Sup problem Had the constants deemed thar cach reo! he pape in {dena dere Sun, Guard, Das pres a. til. the Fina Fl Time we wou hive ba a fowshop problem Aircraft qoeing up cand {san joo, ome tuchine problem, if we wssume, as wed, tat the umber of aiterat srving indy show. Te ea ae the bs ara ‘Me'runway ihe machine. Each aicat hay are ne. namely the (WSaar ne a whch ican gett the apo airapace ane racy aul The abjeeiweheve might betes the werage wating ine of ‘tat before ca and, Pesaps this average shoul be weighed Dy the number of pasengersom och plane. Obvious in eal if ths robles Fmeelsonly part th ar tafe contol efctis. lanes ms ake Moo. Ali igoees the uncertaany inerentimshe sition. ier ‘Ger unpredictable delays ans tnneauer the tme take Land the ocssing tie leach erat ell depend on the Wester eaten of patients a hospital Ayam we gore all andomness Sappose We have patients wo must be {ezned by a consaltane surgeon. The each mes! be M,—seen inthe outpatients department Mirecencd sm the src! wand. prepared for he operation ee Meche operated on Miirecnsr me hp, in he sales] war “Thus each pation (jot) must be proses though eho! our machines Miukte Mand M, above [ee perhaps confsing that n'y one of the {ecentons of prnscsig spot though a "machine seal sig ‘peration, but insehedoling theory all ne operarions Since esc pion ‘ow eley flow trough heuer inthe ordee My My, Ms a ‘Scheduling Problems tha 4s. fomsbop problem, The objective here might be sated a: teat all patente shores possible time, vst gag pony the mos ter schedaliog probes Given these exaples you shoal have no difficuly infin ther pact cal robems ito the general jolr-shop tutute, For insane the flow fre al all this pattern (4) the seedling of ferent programs on 4 computer (@) the processing diferent fatehes of eae ol refi. ia gaape ‘of pat of afferent colours 14 ASSUMPTIONS For the major prt of this book we sl hs making & uber of ast thane abet the sructte of our scheduling problems. Some Were met tioned expliily above; eters were ipl Here we it al the asump- "tons ho fo arthor emphasis and for easy televenee. Why we choose 0 rule thee tsumptions, which are offen gute restive fs disused 1 Each job on enn, Abou the jo is componente opera tions, emo operation of he same jb may be processed simultaneous ‘This we exch rom our dncnon cen prstieal prem. eg those whith components are manufectired simultaneously pir to smembly Int the rte pret 2'No preemption Each operation, once stcte must be completed iofore anther operation nny be startet on that mihine SF Lach oh has m dint operations. one on cach machine. We do not How forthe posslity that jab might requte prosesng twice on the Same machine Equsly, we sass that each Job is prorsted on every ache may not skp one cz owe machines Noe ha his ater con ‘ean sr ilory" Although we coud say tha ajo which skipe ‘chine or proceed upon that machine fr zero tne, we would til Rave S protlem: wher inthe yb posesing sequence shoul I nl oper {lon be placed. Bevaws ‘we dt not low preempeion, the fb coal be elo wating fort machine which not i uct needed 2 Necaelaion, Fach jo mat be proceed to completion, 5. The processing times ane independent 0f the sched In pila we ae ssiming te tings here FHS. e3ch ScCUp tie 6 Sequence Independents sme taken to adjust machine for 9 jb indepen Cental the Job lst prosesed, Secon. the tne 0 moe jb Between, see. 15) Performance Measures ° 6. Improces inventory is allowed, jobs may wai for thei next machine tobe ee. This isnot teal asumpaon. Ia some probs procestng of jobs must be continuous ftom operation to operation. Inset ills oe Ira bas ostke whe the ion ho. 7° Pheri only ane of each ype of machine. We 8 901 allow tht there right bea dhoic of machines nthe procesing of Job. Tht aeumpton "lsmintes, among! oiher, the case where certain Machines have bee, Soplicated to avo Batlenceks Machines may be de No™machine may procas more than One operation ota tne + 10. Machines never Breakdown and are availabe Droghout he scheduling period 11. Te technological conssine oe known in dance ard are mmaable, 12. There temo randomness In parte (2) the numberof jobs is kaon and ae {() the numberof machines kota and xed (©) the processing times are known and tne (G) the ready times are known and fied (©) all other quantities nceded to deine 4 paricular prublem are Xnown and fixed COcesionaly we shall ela one oF two of thete assumptions in specific ‘xamples. When we do, we shal state so exphsty 1.5 PERFORMANCE MEASURES. Tes ot usy’ wo state our abjetins in sche daling. They ae numerous complex and alten coating. Mellor (1966) is 27 Stine schedling onl, lit aa slighty Broader content. tn ation, the mathematics of ‘ur prblcencan be extemay del with even the simplest of objectives. So'me shal not try tobe exnaustve.tntead we shall nae fn general terms afew ofthe cee by which we might sedge ur soe These Ssitenia wl be sutiient co tate the mathematic! definitions ofthe rerformance meanies that ve shal use, Whether these sbtachions a {Imple-minded ha ou theory eases to hava pratal relevance a ovis we should preter wo keup promise delivery date, Otherwie eodwil would strety Be lost ara the might be intel penis a Well ded one may sigue cogenty that a delivery dite that Hot Hem ee ake camping ee ee a 0 ‘Seedling Probes [ont os ‘Sec, 1-5] ene Hee a These quite for pel jb ace redo the Gant agra seunnin Big 12 Theres posh ction fase ofthe tr ilo. Engh lows the oun tn inc cng Tay bo {Sed ote anion ov van tress tad ned Cope ‘So dnc ho cero isin tm: whens proce one, win tinea fw ine refer totem te emarking upon te ambit we hoe tani pou coun. I nyc the next sly estes ‘hat ahe ming lea. We shal offen need Yo dics the mara or the mean of thee describes he flaw pattern or spine within the mache shop. ‘When m = Ia kt Bank. may Be Port Maw-ahup sie, he machine order fora jobs sthe Por the pevmstaion Now.shop ease Here not only the ‘mich onder the sine for ajo, but now we ako ost Uh sath to schadatn sbi the jo oer the ame for fich machine This a sched fr cripletelyspeciied 4 gle permutation ofthe mumbers 2 ‘order tr which the ob are proved on cach and very G. the general jubashop exe where there are 6 restrictions on the form othe technologie consti, Seed a S |B describes the performance measure by which he schedule it be ‘ESfomee. te may lane forte, any of he forms cae i the pesnvs section [Ac an example: m)2/F/}Cy is them job, 2 machine, flow-shop problem here the sm st mats rakes, Tn wag four porameters we follow Conway, Mawel and Nike (1967) Other autors otahy Lensa (1977), Roy Kap (1976), and Gash era (1999), nteduve further parameters bt their dscns ‘atom over anc larger fa of peoblors ham ne shal be oman. 1.7 REAL SCHEDULING PROBL "THE JOR-SHOP ‘One peed aot hae encountered seedling probes in practice w relic that hey ate vais more complen than thos of he jobeshap a3 me ae (etoed them, Few cbey many noch ty uf te asurptions thal me five mal Te est elu schcale slo well cprocnte hy a enc singe Cor. emit be expected heretic haters wo! be ie practical slevance ofthe Thsory which me will evel. Ino, ur hy sty et s examine sumptions 11017 of section 1A. (Assumption 12 ites in moore from the ther apd Will be deuce! separately) Thee aesumpins sne undoubtedly resins. They ht the strstore of the jobshup prviem greats They dete quite recs whch routings of Jaheaze allowable and which ae nt, wha he cups snd avatabites Dt the machines ane ete He istry ey fo imagine pasty occuring ‘oblemse where some of thet resto arc totaly nappropre, So fovnfotiag to fod that these assumptions ate wot actesaty to tbe Aevehoent of a mathematical theory of seedling ator ty ate 9p ‘al of the asunpions which we might make, We could have chown Shomer vt of stamptions, and so defied anater funy of cheduling Problems, aid thes develope theo of cculng fr the fan. Had tre done ne mek! have escomtet he me cle, the sume sme etl we shall ere. The fenom for chovsing the jah fy isthr teats oa preston ofthe theory which patel herent tou Tutors ent ensumbered wins confi of cacy pe ok eto cover speci anes Once the jp la a Beer Stud wi bean eek mater ofall develope fe Moy fe Uaher comtens. To help 0 Hs. Chapier 12 is 8 bret survey of sich Mlevelopment: defines ayer fee chang poh at ‘eterenee the eater appropri 0 tem ‘Astumsion 12 amd, tra smal extent Asmption 1, wand qui /MS AND THE MATHEMATICS OF 6 Scheduling Problem: faut “snes fom te rx, They cone atetion 10 son-apdom problems Teoma aie lf pumeilquamter ae How and Fed in a ro uncon. Benue the numberof jos and thet seats nose ane we calcu problems sate, Heese he Sa pal other pruners ate known anf ell OF at emia robles im shh Job arrive 0.then | Siply pine us that (2-7 (2) x Sl smaller umber, which ponetheless, str getter than 2. Ths x posible fe sae some min Fissera problems, for which tere na anne, no soln. Cleary for Ios peal purges the answer 2000000001 tote above guetion Srstitory, but mathematically quite simply wrong. More inp ‘Sms. and igoosing the nities of mathematical gout. onledge of Uther tere ian exist anewerW a8 optimisation problem in mpor= fant factor ia determining ot spproseh to If we pow there fan ‘pti sluton then we an search fos drety andre importa ‘Suan candidate solution for optinaliy, Suppose, however, we KaDw that for any solution we san sways fd one Bete. albeit oly. marginally fewer Then ne mast adopt a more pragmatic approach, which secs a lation opon which ty inspronement though posniie light. We Seedling Problems [ont Red Br Paper ied Sed wr a © fo fe Ae s 2 34 corer 2 Optimality of Schedules ———————— {24 INTRODUCTION ‘Our problem in scheduling i to timetable the processing of jobs by tachines so thi some caste of performance seheves its optimal value Homever, before ne sporoach this Grea, there are two pertinent ques the wich shouldbe answered, (1) For any particular mesure of performance docs an optimal sehedule cx? (2) Given thet we have foun schedule tats pti with respect ne measure, ow does it fre agaist a sesond mesure? “The fist question ey Dea ite pemplesing unless you ae vent enough Inthe langenge of mathematics to be amare of gfeatest lower Bours [inna compat set, te We shal aod such ers, butte eas ate 38 Foons Supp Lake you to find the smalst umber sry grestet than ile houphtshows that this apparel simple tasks pile eyo give me the umber 2, then U reply that 2 sats greater than 2's gine me slightly greater number say (2 + «) whese¢> 0.then | Siply pine us that (2-7 (2) x Sl smaller umber, which ponetheless, str getter than 2. Ths x posible fe sae some min Fissera problems, for which tere na anne, no soln. Cleary for Ios peal purges the answer 2000000001 tote above guetion Srstitory, but mathematically quite simply wrong. More inp ‘Sms. and igoosing the nities of mathematical gout. onledge of Uther tere ian exist anewerW a8 optimisation problem in mpor= fant factor ia determining ot spproseh to If we pow there fan ‘pti sluton then we an search fos drety andre importa ‘Suan candidate solution for optinaliy, Suppose, however, we KaDw that for any solution we san sways fd one Bete. albeit oly. marginally fewer Then ne mast adopt a more pragmatic approach, which secs a lation opon which ty inspronement though posniie light. We as Optimality of Scteden (Ch? ll an ne ex sexton hat theese oF ep mses sia STheutes do indeed xt Hee ae cond question cle. hough perhars ot ae ae Mle ne mit papas He hee 08 8 Terhune tn seis he pevovoancs esas #5 rae cia hangs doe have to fhe work loner ap? Te stone aninspretain, in secion 11 Tare oF aaa een tba ue aout inertedeows He ACEC! Ur rane eames as toperate fh fal coats hed in rere ea menare pret aly a cmiponent ofthe Lea eran Touro sir a mamnng ne component ce init ede tht minnie ver Se coy re nay ore ssc} than hve het Sen or scan we negate esque Fr 112 REGULAR MEASURES AND SEMLACTIVE SCHEDULES shale’ ant w par of Sn Cher | me o aha T sed the terms “Sequence Nin mote rte mechan or Bi ion Sere wa pay to make me nin between Se 1 GPS ats peccnongseqeace simply he oe whi bare sate rug he tacnee: A proses Sauetic,PeeTOR Om aoe formation shoe! nes a whch he vcs OP i es taan Tale 2 geese of Faces eae. san edu doc. oncver. conan imeabing. wll 3 3 fete cmt, Ths te Gat lage nF. .h peiicn om aeeeeeE tue co ck png the ata rh ine «Pour Pate ee isting nthe process whereby we deve che OM 2 ‘Sequence swing ht an pei sha sas cri pons a approach nil neo folios. Fy we shill ot hs tere a a Se sree lace poceinasegveces. Then we sl show that the RRO chuetporrmance messes we ned only conser one frm of se atace thos once we epee tus hs nd nl his titahing, 22 cane dsc aunque endl I fle tat we oy fae inte Seaeae Yes to Consider, Now i ti tenes ta ema the meme oan opm! schedule. Inking yo fo fed th sales el cee ly grote han 2,1 was aking 30 0 seach a OC See eritiy sua numbers sghy arene tan 2 Such a search er Sodan But an earch of ie set mus eventually end. Here ran aetna of eed, Soto nde wth he sea value of Toe datatce eeasure all we need dof scorch rough a he po Sec 2.2] Regular Mesures and Semi-Active Schedubes, 2 hmaue andthe smlest hrs bot found. A lest, comceproally nat all uetneed dete pate the finite mumber of sched ny be it [ret tor such um exhaustive senrch oe humanly pss See the wl $n wo Prem 1.2.2 in secon 8) However, only the conceptual fry tat we mca orc to sow that an optimal schedule cis Sor m hep by noting thee ae only m Gait numberof preg sequen I fact we have sready shown this in out shton Problem 1133" Sell wl do no acm 10 repeat the argument bere A prowess sequeree ives for each machine the onler of prcesing jobs Thus i Species obe permatatiin of HJ uJ.) fr ash of he or aches, Shes tere aren! permatatons of yl J) there ae (a!) pose Feovesingorder. Actual, Because! he tchnologel costa, many ‘EF ihcse moy Be tlie. Than thre may Beles than (1 process ‘SEquence ts Scie ut, whatever the ean, cour ha he dumber Timetalg ced semiactve tin the ating eho there does omen vt ih cn ee cr et aie peveeung sequence 07 vain the technolopeal concn and 40) Hac in nice ond, in scmbscivetimeabng the pros tt Spear at wom an ean be yom ok bak sp eS Seas sou wl se that there be wed seme neta wth Saesoning why. jst cerned the sruble thing w do. However we Shovld queso ha tation Specialy sould Algy nal bi re awe fe oh county cae we had asencd pss mee {Ghar hod mot picked up the Gara a 1000 am, {ne inate? Srey nt et we a uy hi But hfe for “Theorem 21. oder to minimis reglarmessure of pexfomancs is tony neces 0 consider sem-active tnetbling Proof Consider a schedule which has not een constr by semiacive Uimetabiag, Then there st Ket one operation which souk fe stared euties, Ot al ich operations chore one with the eae fishing ine Reumetae this operation to stat a carl aw posible. th ths. pew schedule here cannot be 3 job Whos compton ime is incesse. Ts the vale of a rogult measure cimnct have meresed etc. Reps this proces of retinetabing operations, which could have seared ear, ui thee ate mons such Teft Note ta, becuse ne aby Pick an operation which has the earest completion tne. no operation { reimetables more than once. There ares finite numberof opens, 50 his setinealingprosest st terme The final Sched the oul of sem-acive tmetabig, since no operation could be sated ‘ali Moreover. since the vale f the performance mcesrs does ot By ‘Optimality of Schedules (2 snereaeatuny sing, th ia chee a eat as g00d 25 she igi en Taglar messueee perrmance seas tmetabling poduees aa ean pod a htae weight find by anyother meth net afin he ate poo oul you may benefit by daw ing our Gane gen wit sig sare imei 8 then retinetabiag operons as eset nts ns Pie of Algy an fen we wv 0 mT gic a oar MELE ‘Noein wep our im argent sould Be leat, sty 2 Fen ee paces eywenss. Por cs possi tnt ere incite pees aon dang That Fe et number ofthese schedules. Meorem 2.1 es fat ae un pertormance eure we need cet 10 OMe ‘Toes asc ew maiming fonction oe fie a2 se gud eons. There my. of comse, Be more than me ial she fe ec, tun breiy fetur oor sine Be er mcs and sues. We shall rms now on confine aa ees emarly to reguar measures and, as 2 comequencs of ametee Sspiy comer heduls consraced by sembacive tine Theorem ohne tu asequeace unucy defines the nocd sched ane eunally voting to be ened by pedanaiycomtinaing 1 Bancion ond me shall aot do 2.3 RELATIONS BETWEEN PERFORMANCE MEASURES ‘We shai uy tht two poformance measures are equate» sched aie eal ah espet fo one hls cptial ah respct 10 te oe eran setom we shal prove simple theorems show: a eo he ease trosoced in Chapter 1 are equivalent, Let tbe withthe seme ‘Theoem 2.2, The flowing peformance messes ar equiva we WR GM OE rot From the dfiitons in sestion 1.5 we have for each job 3, the Felon ee especialy Fig, 1.2) Lea, aa) Summing ovr the jobs and viding By mes the relations between the Gant WS rea quate: CaF ere W + Cin®) BrpereLra ey ‘Now the quantities? (In 3) 327) andi are constants foreach prot seule] en a =e lem and indpendt ofthe shade Tha in civouing 4 ttecle 10 ammie sea ato mininiing PW. and Sima) sua Palen sumntes and and wo on The fow ese ae equent incr, minmsing he mea compton time of he fb so la natn rman ote meaning tine mdmean cess Wesbuls fom ete there hr pete es cone ring Fre We ad our meas ‘euialent Inte above peo the top ‘ltcom 2.1) and (2) valid whom we take the mean of the quantities adv. ut not wen we Take the maximum. For instance, sot Erealp ish Fant Fo Come ace pn Bin =O pu eSs Me IO pa Vig 21 shows the schedule whith processes J then J Here we hase C2562 ao Cea BaSB= 1 bam 5, ne0 n= 10 r= and Hence me Cay ~ U4 $410 ~ Fag + og OF coun there are Special cacy whem ran of the encaares Com Pe, Wore ned Eo ae ‘uivalent. fn problems with al ready tines 29 C,-yand Fare equal rt (hecuse €,~ Ff al obs) Again. problern bere all the de ‘dis ase equi 0 some constant dC And gy ne eval (boca Cand Liter bythe same constant d fr al fobs). Ter us now nae the partial equalence uf Land Ta. ‘Teor 23. A schedule which i opal With spt 10 Dis a0 wth eapet to os Proof Ty ~ mas|mast lt, ma 0), manly 2,0) merited) maxi.) So mipsnsing fy ako mises Tr NIB We dana cla tat MOINS Toy RIES Lge Any sehedule which 'nshes all the jobe on of Bete their dus dates as 30 “Optima of Schedules fen? 1 be otter Tax=0, #8 minim posse vale, However there Sous roy ‘put leaving Togo en ae fo even ease, 0 tM La Tae next theorem confirms what sears intuitively reasonable. Namely ree ae al sumption ne fal the obs en he YORE A ina eer beng used stay one tine © nasi an ime lee of machine eines Teor 24, The flowing messares ae even DCm GN, GOT (NB. We sek o minimise Can and. but 9 maximise 8) ‘th equivalence of @) and (i). secondly the oe ca i) ten FON Proof, We sal show f Proeh sone nf) an (i) he equivalence of (i) and i imei Equivalence of 6) sed (i 1 Rya ef *Me de by detiion need Me iow Ny) ste aun of nbs atl Being proceed me 1 at tac bing poesedon# s-{ actin ie Lo otherwise. ne nyo Sa so np CH) #1 opty Be aes [sande = S p,- mie [Tae 87H 3 Sp = ze ex W ising Coa ie ani fons that maximising Fs equivalent 16 iimsing Co Bip, constant independent ofthe shee See. 24) References and Further Reading au guvaence of () ad ih Note fs that, = Cay ~ 2 py: that hele time on Ms he di {me betmeen the total schedle length ad the Toa proses time > foc smacking, Tas 7 1 Sa -Lintn-$ $00 igo. Cy and iter bya constontindepent ofthe hed, am hence it {elon fat hnimsing Cao minimises and vic ers Remembering a rina replat messi We sy Sede shat the cauivslet measre Fs ad reper Simone ight once that Ny ‘eplar ad, bose spesking this sw. However. to be more presse we ‘ould eal that nese to minim o regular measure and se pcan tere since ne win o mcm We may svercne tea yoy te icon of ius sm Scking to minimise (-X,) pec Theorem 2.4 that (-R,) fren meas. : say ein erm sg and Ry ht ot Toeorem 2.5. and CC A eguvaent mciures of perf ‘Theorem 2.6. Ny and WiC ave equivalent messes of performance In geoera Z/C 08 Cy, ae mot elt meres of performance. (See Problem 237) Turn med ede tat ithe Nor Anregae Honever te soc cn fa single wach pele we toy aint reguay: Whur proseing nue We Coe ‘Sraantmimel the han fhe proc nes, Hes ov ing ‘thn som’ cea any eqn 8 Comb Corry 27. For snge machine problems, the follwing masues ae ‘equivalent: a OC. GFL GH GL. OI RL WH Re 2:4 REERRENCES AND FURTHER READING Coma. saxell and Miler (1967, pp. 9-21. 9p. 109-112) nd Rinaooy Kan (1976. pp. 16-24) mere the pritial sours of eelerence fr the shove. The concept of 9 emitting schedule ray be further veloped a Optimality of Schedules jon? int that ofan active shel, We shall dig us his in Chapter 10. ew ‘Sinors seem to have ebasdered the equvalene of performance mea raet pe they have concentrated her efforts o slg paricular prob ESE Apr tron te eo hooks et wbore, the ony referee of mote {bat of Gupta and Desc (1071). where te posible equivalence of cer 12a Peseremance mcamures and te ttl seedling ost are investigated ‘epi 2.5 PROBLEMS 1. For weight ayes +8, Sw Matte flowing meses ae @ Sate wy Sew, 0) Sat w Ser 2. Define the met ust of he machines ote average propo sion ofthe mike-spen for whih the machines are actually processing Roe hat using mean winston we equivalent 19 minimising make san sf "show tha igimising a meightes average ile time vr. 2 a lihere sp and 3 fos equivalent © minimising the unweighted sserageT Phir a 2 job, L machine example with procesing times 3 and respects an de dates tad 4 respectively Blah are sea for proces SAPS Ginc aero Evaluate and Tf te two pose sheds and se dhace hata schedule abi f optimal wih vest to. eed not 1 optimal wih espect to T= nor ie ers Powe Theorem 25 f Pve Theorem 26 $s Ecos not regular measur of pesformance ty consi ving the 22)G(C/C exam with Sata ‘Tecnologia Constr ae Processing Ties Joh st Sache 2 Machine yo Mepast | Micpurt 1 Mopae Mepis Both jobs are ready foe processing imma see 251 Probleme 3 ns Compare the seeds Machine Praceaing Sequence My Low M. (6) Show H/C 4 08» repelar measure ot performance by consider ne SiGre) example wih 4 Tectolopeal Contrainte and Procasog Ties bb IstMachine 2nd Machine 3rd Machine Miter 10 Mapes! Myspana Maps Mupas 1 Mupa=9 i mila gees ee he tac jobs are ready for procesing immediate How: Compare the chctles ° Mtoe —ProceingSequene Machine Proeing Sequence M Boh kM i eo s Rok Rom fee a Boe OR Mm Bo OR Chaps 3 Single Machine Processing: Basic Results 341 INTRODUCTION We are now in postion to begin solving problems, adaitely no part lary ccut one as this chapter soni concerned with the very sp ‘of those volving one machine alone. Nonetheles, we stop claw discussing exitence and equivalence of solutions and actually attempt salve someting "We shall assume throughout this chapter that all obs are ready oe processing atthe beginning ofthe processing period. Perhaps it would be test emphasise oth this and the fact tat we are ony considering sing machin problems Assumpdons fr Chapter 3 no forall J, (man Fen m (General in each chapter we shall make a number of assumptions over and above thoe listed in econ 1.4. These wl py forthe dation of ‘hat chaper alone and we salt dem explcil,ahere, nthe itrodue ton) tis ieresting to ote that single machin scheduling problems aria practice more often that one might expect Fist, tere are obvious once Involving single machine, eg. the processing of bs trough a small nomtime-sharing computer Thea, tere ae ls obvious ones where a large compiex plant ats asi were one machine in pat manufac: tare the whol lant may ave oe devoted to making One colour of paint ‘atime. Finally there are jobshops with more than one machine, bat in which one machine acts asa ‘borde-aec’ inthe hospital example of Section 13 the treatment of patents may be severely restricted by the Shortage of theate faites, Thus M, the actual rrp operation, ey sec 3:2] Permutation Schedules Ey {42 PERMUTATION SCHEDULES Ja the example of Alay and his frends we sw that certain schedules, a lor the one in Table 1S, ease the machines, the papers to be {ic when perhaps they ned not be. For instance Algy could hve started fovend the nancial Tones at 8.30, Dat he leaves ile 0 that Charles fay have it fist. Ths ace, not procesing the obs immediatly avalable, bet wating for on that wl shorty be 30, maybe employed by the optimal Schedule the general nV problem However, me sal ow se hat iris unnecessary to comsider such inserted dle time fora single machine problem with a ropular measure of performance In fat, ou rest fllos amesiatl fom Tacorem 2.1 (Why?) Nonetheless, thee is po harm proving i afresh from frst principe and doing 0 wll emphasis ote of rpument tht i common fo much of our subject, ‘eorem 3.1 For a8 n/1//B problem. where B is a regular measure of peformance there exists an optimal schedule in Which thee so lserted {etme ie. the machine stats processing at~ O and continues without fest ntl f= Crue Poof By he mgunent insect 22 we know an opin hades (Si egunr mesure) Lerten heute ae teem tmariossna, keene ie seeteobae Fo sy ‘eis fcend ia amonar (pseu opeatoneecemr ae spre tar Sh fee ance tee banlyone hachneeCeayaie ‘enicton umes ders are ch Gx Gofoe= de \wtere the Care the completion times under S: Thus the vale of any ‘par measure cannot increase in passing from 5 t0 8 SoS must he ‘ptinal too: This process maybe repeated to remove all asered pevodsof ‘ie time, so leaving an optimal schedule in which processing coniucus. In section 14 we made the assumption that operations could not be Dreempied; ie. we are not allowed to interrupt tbe processing of one ‘peation bythe processing of seond before returning to complet that ‘theft For the general n/n problem thi consent on out ae ‘Single Machine Processing: Basic Results: (ag soltion; we might do beter wih sched invaing pre-emption 1 ‘ecm do withthe best oe tat doe mr Hoeven gle mache {Sse with 3 regular measure of pertormance tis snot wy there no vantage tobe sind ining preemption, hor 22- Te an//B ble, wre Bi pl mene of ee formance. no ampronemen may te pind nthe opal schecle lowing pre-crpton q Prof (Sketch, ee Probie 3.7.1). Suppo weave aa optimal achedle Sibi ob eps so allo job io ‘{ibostocecmptonstsomelatrsing Sethe Cant dap ni 3 cassume K comptes nihout pre-emption sae ™be-twcedle sine “nrc tne eh tle cetven a need eee ee om erate aa TT Searls fe apr eae ‘there exists am optumal schedule entirely without pre-emption. 7 Th ptf lg ut otal aera teers pls ern de lpemep, cuebinny ottependincanheer neal ee cence orrie ree ee cpa gts ob de om oe ate eee) andes Be ob ht te procengsguence set athe ih be pred, se 7 i oo baa @ 431 pam aS see 3.8) ‘Shortest Processing Time Scheduling, = 4.3 SHORTEST PROCESSING TIME SCHEDULING SPT mia F, 2, i) 1,1, ‘Suppose that we wih 1 minimise mean fw time, Le. our problem is an Foe prtcalr rocesing Sequence orn) (onordering he sum ain ‘the processing sequence) ow Si.spyy = 2p ita contat forall sequences. Hence minimise Fac nimsc 8 Way we chose a equence toma cach Wy ‘Sl cul pouiy Be, then we thal lai mine the um Wan forall queens sats media Wi) pyorine emt watony or jt be proesed. Thu ite cose’ "to have the shores procesing tie ofall the jobe ede) we Sal mise Way Wy pes ion amc fam wat oly for J and Jy © be Pro= cued "tna mise Wyyte chow yy and Jy have the sore taoext hott posing ne tom he ty JJ) weet i Si hve te shorten then we donot afl ur cai maton of Ios Therefore we cn mimi Wy ad Wp silane “Eontinsing hs ay we bud up schedule which a the th job 0 be rowed far the shoves prog tine of on remanng Dag “Sutacoualy minimises al'be wang ns. Thus the rey an SPT Seindale (Soret Procning Tine). minimises F. We sate ‘hs a ‘Ticorem $3; 0 which we sn provide tn lease poo. The reson or ‘roving hi thre again tat the strate prot has a orm common {nan of our ter argument 1 doesnot have the abranige, however, ‘ing ely cnstuctve a he poo bow ‘Theorem 3.3. Foran n/(/F problem, the mean flow ime is minimised by sequencing such that 2.) Pax Pao + $Paoe nore pais the processing time ofthe Job that is proces th Proof (Alternative), Let bea non SPT schedule. Then fr some Pay > Pa on Let be the schedule obtained by interchanging Ja ad Jaa S (8 coal ‘Stale Machine Processing: Basic Results = [Ch.3 Fig. 3.2). For comenience, label J 8 and J. a8 K. Albjobe others than and have te same ow tanonsm San in. So the diference i tein tine oS an" depen oom te Now tines foros an So [Mein Let cose Are wins’ Soins, Fratp and Fane tpt pe 02) Similarly ins’, Fi=a rpc +p) and Fea py oa ens the obo of Fan to Fin BOMB -heesn+ntn ee = huis, thei contibusion 0 in “Uns has smaler meon flow time than and so any now SE scnedae can be boner. Inerefne an SPL senede mst soe an dF Hublem-(N.B. If two or more Jobs have cgual processing es, here Wl Ie eore than one SPT schedule) From the equivalene of Coroliry 2.7 we may also ote that SPT scheduling solves the following problems yy, mi, afl, ni mA ‘als minimis the mean completion time, the mean wating ime, he mean latenes the mean numberof wnfinshed jobs, and the med number OF jobs wating Between machines, oh pee Le Fis athite eee before fs —4 tht et 2 es © et fee Bt Icy eb emily dB eae es EBT ‘see. 34) Earliest Due Date Scheduling » 2f ale. Ginn he fling 7/1 poem J prsases ProceningTime 6 4 8 3°27 1 we fi hat the opimal SPT schedule is (7 5,4, 2.1.6.3) be do job 7 then ob 5 and soon Tovcalculate the optimal vale of F we note that +8 oH TE py a Wore 624834443 6r 27TH) = 4x90 = 124 Generally we note tha for the SPT schedule Fat Sk Dw oo WAZ) oy 24 EARLIEST DUE DATE SCHEDULING EDD nue min Leow hT An intial approach to scheduling is perhaps. to sequence absinthe order invwhich they are required. In other words, to sequence the jobs such hat the fist procesed has cares due date, the second prosesed the net ‘arlst duo due. and soon. What does this accompli? The answer ven by ‘Theorem SA, FoF 30/1) problem, the maximum fatness i ini red by sequencing such hat tay = where da the de date ofthe ob hat is processed kth Proof, Suppose Sica schedule in which he js te not ordered according, to increasing die-ate. Then for some fy dyy > dy, FOr conVERIENCY, ube jobs J a0 Jn 88 Fad Krespetely. So we have sts aod Shek 39) tte ES one ze ke dae Ot fede dee at fetfends bin ers oe ‘Single Machine Processing: Basic Results [a3 itimay hep to look back 1 ig 32 in tte peot orien 33) oreeeree Tet be the maximum lateness ofthe (a = 2) jobs oer than Tan K under S and” be the corresponding quantity ander 8" Clesly [et Ly Ly be the lateness of and Kunet Sand Ey be the corespond ing quanties under We, therefor, have the matmim atcnen er S Lan = met(LLp by ‘nd the maxim ltenes onder Ling = man's Li) man(L tbh since b= 1 Now underS Li + Cr ~de atthe : Le FAT Be— dys MA” BP and unite Lina teconnd Terre etre (0:9, > 0, wad te>li OOS) Hen a> mal onatiatnidcnatting Sst) eeararitn) So ere By Theorem 2.3 we may note that we have also solved The 1/1/Tra, prot. This mth of sequencing eles Ear hen ‘cariest omitted. o peaeeeae ‘Example. Suppose that we have to sve the n/a problem Job 123 456 Due Due Tse Res Procewing Time 1102 4 7 3 Sec. 3.5] ‘Mecee's Algerian a ‘teal a9 optimal EDD sequence is (6 2,1, 3.5, 491; We may caealate {bc oplinal Fan though a tabular form of calulation as below: Here we find Fes 1 Job Compejon Time Lateness Tardoes 3a Pan Gan = 3 Pan Ais) Lay Candy Tas ™ MAROC) a3 > 0 ° 2 4 sii: 1 eee t2 ° 327 = 6 sta ee ° 442 ao 6 NB. To form the Completion Time column we simply ad the pressing {ie ofthe cucent od fo he completion ime ofthe oe preceding. Thos “ihas processing tne 1, Jy completed at = 7-So J, completey at “his mean that nen of (=I) addons to form the sam Ef 10 {he th row ofthe able, thee is nly one. We shall we this il ober ‘ation in Chapter 6 when we sal eed to calculate T for shed 35 MOORE'S ALGORITHM. aig My Ase have sugested in section 15, it occasionally makes sense to penalise ‘andy jo ually however late hey ae. Thus it would ost just a8 much (Gray a doe date by one week at by one hundred yats. For instance, Fier and aura 1978) we hind of pea bli sir ‘ute Tights since to miss launch date even by a fw hours can com> Plesiy ape & space min. If we adopt this phlowphy, our scheduling fost spb ommimise the umber of a problem, We conser en algorithm for saving ths due to Moore, but in 3 oem suggested by Hodgson. We fist sate the algorithm and give an ‘ample ofits use. Then we shall prove that does indeed find an optimal ‘hed, Algorithm 3.5 (Moore and Hades) Sep I Sequence the jobs assordng 10 the EDD rule 1 find the current, sequence On Deo) SEH ta Aq) dyey for - ‘Single Machine Processing: Basie Results = (Ch. 3. Sep 2: Find the fst ard ob, ay Boy the current sequence Ifo such jo 6 Found, go 10 Sup ‘Step 5: Find the jo in the seauence (ay Byy J) with the Largest processing time and rejects fm the coren sequence. Return to Srp 2 vith a curren sequcnes on shorter than before, Suep & Form an optimal schedule by taking the curent sequence and appending tothe rejected jobs, wh nay be sequenced in ay ore NB. The cece jobs wil all be tardy and these wil e the only tardy jobs Example. Consider the 6///aproben: sab aise Dae Dae 13 6 9 3 20 30 Prcening Tine 10 3 4 8 10 6 ‘We fst orm the EDD sequence and compute the completion times until tardy job found (Step I and 2). Curren Sequence 23 1 8 4 6 Due Date 6 9 18 2 2 30 Procesing Time (3 4 (0) 10 8 6 Completon Time 37 17 Job 1 isthe first tardy job in the sequence and of the subsequence (2.3.1) ‘thas the largest processing tine. So we reject Job I (Step 3), We return 4nd repeat Step 2 withthe now curtent sequence New Current Sequence 2 3S 4 6 Rejected Join Dus Dae 69 0 23 30 0 Procesing Time = (3. 4 0) 8) 6 Completon Time 37-17 a5 Job 4 isthe fit tardy job in this sequence and of the subsequence (2.3,5)') Job 5 asthe largest procesing time So we reject it (Step 3). Retuming to Step 2, me fn! there are no frter tardy jobs New Curreat Sequence 2 3 4 6 Rejected Jobe Dee Due é Processing Tame 3 Completin Time 3 2 1s 28 ‘See. 3.5] ‘Moore's Algorithm: a Hence we pas to Step 4 and form the optimal sequence (23:4, 61.5. Note tot (23, 4,6 3,1) alo optimal. We can ly ut the ealelation mote compact frm EDD Sequence 2-31 S66 Rejected Jobe Due Date eos 0 8 Procening Time 3 4 10 10 8S. Completion Tine 3717 o CompetonTine 37° 1728S Completion Time 27-15 2 ln forming the above table there is one Completion Tie row for exch {ele ofthe algorithm, When a jb rejected, its murber noted in the ‘gt hand column and its completion tne i ace ccs simply blake ot with am asterisk, ‘We now show that Moore and Hodgson’ algorithm does produce an ‘optimal sedule Fist ote tat any schedule that produces asthe form Se= Awe) whore Ay is the sequence of jobs sompleed on time and ordered by the oad Ry Sethe arbitrary sequence of the je that are tardy under Sy ‘ur it tp isto show that any optimal schedule can be taken to have 3 ‘irlarformS= (AR, atbouph we do ot clan thatthe jos in A are necessary the Same as those it Ay Le be any optimal schedule and R the set of jobs tardy ander S, Resequence so thatthe jobs in are scheduled lat ana the order of he remaining job is uachanged. Let S"~ (B,) be the seauence so formed, Nate that the jbs in Bare arted no later in than they were in 8. Ths Po Jobin B can be tardy and the numberof jobs tardy under Sino treater than under S. Next observe that, nce there ae n tardy obs in B, K=O for this subsequence. Since the EDD role minimises Las (icorem 3.4), esequencng according to EDD doesnot increase Lon “Thus, i is the vest of resequencing Bn this way, no Job scheduled cording to 4 can haye pose latenes, 2. can be tardy. So the number ‘t jbs tary under S" = (A,R) i 9 greater than unde 5 8 optimal, "hs sao. Hence we have an optimal schedule ofthe required form. “The nex stage of eu" pool i o show thatthe numberof tate jhe tunder'a schedule produced by Moore and Hodgins ait 90 “4 ‘Single Machine Procesing: Basi: Results [Ch.3 “eater than tha under anv optim sede Set Se (Au) bea schedule prodced by Moots algorithm, 5 (AIR) be an optimal schedule, and 7) be the numberof obs ate under Sy wal elp our aoaton to renunber al the jb 0 tha (J.-A in EDD onde, Le-so that Bedi for £12 cos D “This Ut.) isthe sequence of jobs proced by Step 1 of the got, Let Jy Jy yr Jy Be he sequence of ext at jobs found in tgp 2. Let Jey. Jybe the sequence of he bs rejected by Step 3. Thus Jy i the Jo tha Fjete in the wh eyele on dncove ing that Jy isthe next ate jb inthe curent Sequence. Te-sowe Sys “optima semewearermwunder SiS. We need one more piece af notation. Let be the umber of jabs (Jody) tat ae ary wader SEE for 1.2." and enue tha he womb of) ary «how at “x “ner Sh rater than tha ude. onion nico sin tosehedte Be PDN De dey Jy) bt at ea oe beg ad The ee ef i apis! iy needs wale ‘Ae ote hat he jo rejected by mame Jyy asthe lng proses Sng time ofthe sequence UJ» Pai taj mi fhe by Js ore jn reese by 5 which haste loge procesig tie “has immedi hat y= Puy ee Jo Pelee By Sy as 8 Procening tne es at a8 longest of the rected YS. ‘Semamumestaiocsomesaet we hve show tat wet (B) Joop Jy 628 be pied with jobs Iyoy Jyay <> Syn i ‘hey lected by such tat Pan ® Paw) FOE =D ok Weshattsnow (Ayan (BYE WOID er), Since we have shown they sre uve fort = 1, we can now cli these relation togeter to show By induction that A and B hold for t= Tes wal hat pa SO, > EWE RIVE fs 4 Te Ge 5, “Thus assume’ p,~ | By our pang we have thatthe jobe rejected under Sy and we such that Spm Span Spear Thos eae many oetler than te abternle § ‘Sec. 36] References and Further Reading. as ‘hus he “time saved’ by the rejections unde a Het vider S- Hence unde 3,085 Oey Jyoy- = Jyo) stat ltr tha they o under S By constton of. we adw one ofthese jobs must bb tandy under Sy 20 a leet One mus ao be late under S-Thus a least, thc hese obs ma be rejected by Sandy, must therefore be atleast, ‘no pteater than SO po Pw PE SES Tan requiced, ‘omen (A) and (8) we have show 4 Lhe (A bl fe We ‘Sos now show tacshe pain candnion (Bhan haafort = 4 To do {hee simply pate job J withthe longest unpaired job J. ofthe dn sobsrejeted by S-Because J. ihe longest jobs ar united by Sein the sequence by Ja sss yy. and beats yy Joys Jon Bake fifoen pied wit the longest then unpaired jobs mong thos eected by Se me ee that Pay = Py 8 eae. "We have now all bu shed, We have shown: pes = me? S intum, f2 8 2, finally, 4, = “Thos eeecs at east as many jo 8 Sy 5 opimal therefore Sy is ab. Weave proved ‘Theorem 3.6, Algorithm 35 finds an optimal schedule for an n/fny problem, |\6 REFERENCES AND FURTHER READING ‘Conway, Maxwell and Miler (1967) and Baker (1974) deal with single tmacnine processing fn considerable deta nd indeed we shall seus he Topi further oureives, principally ip Chapters and 6. Rianooy Kan [UB%6) and Grahame af (1979) survey the majority of he results in the “The SPT sequencing result very oki and has been proved na number of ways: Conway, Maxell and Miler (1967p. 26 give am ntersting and Insrctve proof based upon the aea under x concave graph and Rinncoy Kan (1976p. 88) provides an algebra pro ite dine from either of tur above. "The EDD sequencing rest was originally due to Jackson (953), "The development of Moore and Hodpwon's algo as orgioally {ren by Moore (1968) nas mach longer than that ge above. urs i ‘osc pon one of Sturm’ (1970). Kise ea. (1978) bave modiied and ‘extended this algorithm so that it solves n/n problems in which there 46 ‘Single Machine Processing: Basie Results [Ch.3 {3:6 non-zero ready ies, but withthe rete that d= F< (be the teady times must incteae in the same sequences the die dacs. Sidney (1973) hac modied Moore's result in diferent deci He algorithm allows that certain obs be spesied as on-tary.e the proce sing sequence must ensure hat these ae completed efor hc de dates, 37 PROBLEMS ZA Complete 308 proot of Theorem 32 by: (2) allowing that job T might be Bren ito more than two segments by pre-emption; and (6) allowing hat the jo K may tal be pre-empted. 7 2. Usethe method of the ateraive prot of Theorem 33 1 show SPT scheduling solves the n//(n 37°) problem ie a mini ies mean square fom time, 7-3 Prove formula (3.4) above and find similar expesion for Z |. & Solve the n/Y/C problem below and forthe optimal schedle fn OWN, sb ras4asere Procesing Tine 10 247 313 2 5. You ate asked to solve the problem n///ZaF, where S) land > Ofori = 1.2... ms. you hive to miione 4 weighed ean ‘of the Hw times. Show tht optimal schedule sequence he jobs such tat ‘What would the optinal sequence be if some of the 2, = 0? Solve the I problem and find the optimal J Tea aecaeer a Due Date 80 20 67 48 1m 30 0 ProesingTime 29°13 31 20 7 3 9 7 Atany time define the slack time of an unprocessed job 3, 38 2, = 1). 4e. the amount of time temining before must be sated it ‘order that I tished on time. Consigr sequencing according 10 nom decreasing ack time: In other words, when one job completes elect the ‘ext fo be the one which has the shortest slack tne. Show tha thi se. ‘quence maximises the minimum job lateness ‘see. 3.7) (Problems a Solve the 10/1; problem soo Pouoie sere pe +b 6 Baeh NB 2 PomumTine $3 124 4 2°11 4 7 9. NASA. has one space shutle with which to launch 8 spaces tions, Each station specifically designed to perform certain astonomea ‘Spicrvatons Each nt be placed in poston by aertnin date ort nl be to. Given the data below, in what order shold the stations be Tmunshed? Assume thatthe Lauaching and construction sequence saris ftom Ist January 1982, “Time toad nto shutl, aunch, Mat ein Suaon anc bil In pace mie 7 1 year, 2 mone ‘ape 986 2 month Br is 3 At monte ‘ve 1983 4 3 mons Sar 1986 3 onthe Sep 988 months hay 1982 5 mone Det tone : 1 yer, 2 ome Soe 1984 10, Conder an n//Lne poem n which he ready ines are nt secomniy ob ntfs m whch he proses ies re Shap eal p= 1) Comte a agosto for dg 80 ota Shee Fi Note ht, sie rt tal ad hep, fend arg the poceing of soe, no job can become Chapter 4 Single Machine Processing: Precedence Constraints and Efficiency 4.1 INTRODUCTION Fw iam coms all thei esters 40 be of saul importance. Foe sample, sappone thi Here a numberof obs fe peo a hat tne fora teslomer of exceptional imprtanc,ehowe gobi ust Be ‘taeda all cox tnt an, the frm te exremels ely Seode that ns job sould Re shed thragh nd lesahed tet whatever the Eonsegdenes fr the seeing of he ater jobs Hers we have ah exam Pie ofa seedling protien uhh i complies bythe exec of pre- {edence constrains In general the mtd choxe ofscheduleby demans tnany ox racic seedling problems so which precedence constrains fist ll esufiient for our purponc cole just two fer tere Fist cosidr situation where any adhsament to 2 eachie' st tings may requite a substan inc wo ake effect pote certain eye hems need to arm tp or cost down In sich cane H mes sense process topthcr ay groupe oh that euine sii settings boca the hange-ver tines tween the jon wl be slight. Second comer the ‘chedaling of program po 2 mute. Seppe that one program Pe Ahuces an output fe tats secu vegies xp ta Ten oN these me casi hat te ate uly eased before the secon begun. tn Both examples thereat clear prvedemes comin pon the ‘Seating the jobs aysover satu i ay at Sst oem to model certain aspcets of 8 scnedling problem using pocedsnce coma soften eter heorel taly wo we other deyaces Heal mporant customers an simla e Pisstom shoul! be treated by an appropmate choice wt perormance Ineasre ome thit pees a stedule very heavy 8 parts 8 fard, Sima. when the stp tne stn pation depends rota ton whih job ste peated ramus Bern. me should bot nr Ace procedence coment nt our mode, Rather we Soul snd ‘See 41] EE - Assumption $ of section | snd admit inte tit the procesing Himes, hic eclage the stp times ane not neceseri scyacnee independent Fmever the coretty formu prablemc may be very cticl fo save sthorss the presence ot pres i ay actully help in the ‘luton problem brcause they reduce the number of faxes {uetor. Hu this does ae mean that you shoul ink of precedence or ‘Hrin asolty a tek hy which we ay avon partly iil pod iene In very many canes they eater a problem guile eatualy Foe inwanee inthe computing example ts or Because of cost hat we seek 0 ran the second propam afr he ist simply Becasc of lag eesti, TW undersiandable that some pooplé should confuse techaoogial swans und precedense contains so we shal pause ad ty 10 make tie dstintion clear Techologa!constaits give te ede 8 WA he ett that sompose each Job aust be processed. In short they Be {He outing tht earh ib mt follow tween the machines, Precedence trains on tie ocr and, rect the sequencs of proscssngopers thonsbetween decent jobs, Tey iat tats eran persion of one jo pun he fully procesed before acetin operation of diferent jo may Be ‘gon Often et hyo mesos alnaye, the reaction applies tothe “peaton ofthe freer job athe st operation ofthe te, so ssing ioc ee ose ja i ompleed toe te others begun, Perhaps the emer tay tw apyueciae te dation between technological and. precedence Commie to otter ingle icine procesing, Becine ther fs ony ine poration per jo. there cam be mo technol contains here however, there ca yuitecealy be precedence constants, each of he thn beimpls quted above can apply he single machine ease Tn thi chap ne comer two spl cas of singe machine prob lean mith recesence constant Tn Seton ft we sonener another sass of ach problems. Final in Chapec 12 we futher escss the form that Presedence consrats ay take anid extensively reference the Ieratare ious of eating tht th the flow shop ad oR shop, 38 ell =m sve general scheduling problems [ater hs chaper We tr fom precedence constraints te ein. This tie cam bet He sna By refering back to oor Sess cota]? There we ddruted hut nur choise of perfomance mesnare S Tnortere the erchange of songs O and beings 8 net detest Inthe ttl flow tre and ene the mea i Tee simp matter to apy th theorem ina probe to dedioe on ‘optial schedule Tas only omack pois wordy raking i hat 9 Jo wth éoes note aby steng soul be teat ven degenerate tng vont co ee ss tamuses atcormin be Apfice ot p s In this seston we dacs an algorithm which was developed by Laver (1973) co deal with rather more general precedence constants Here we ‘hal simply be comsraies to pokes cet jb bore, ut ot mesa as enmedtely before, certain eters. Lamler's agiihms mins the ‘maximum cost of proesing a job. where this cost ha general form 7(C) {or ands taken to Be nom-Sstesing i the complete tine C, Ths the lgortns siimises ge Hon Because of he aon decreasing nature ofeach (it foiows mediately hat th pecormance mesure regular At fast ight (43) looks some ‘tha foreboding compared th out sarher performance measures. bow ve. for specie choices y(C,) (48) takes on much ore fa forms. WC)eGns= Cy then (OS) ger the measure Lays HE) = manic, — 4,0}. ges Tp. To develop the algorithm w2 need the folowing sheorem Theorem 42. Consider the a/t//max/n(€)} problem with precedence onstan Let V denote the sibs of jb which nay he performed I Le those jobs which are not regired 0 precede any other Note that the final job the schedule mus complete at += 3p, Let J.be a job ia seis “9 see 41 "Algorithm Ss 40) ~ pigtde it he ws that my be pertnme st , acurs the ee os. ten ache In hic eee ast prot Suppose $ isan optimal schedule with not ao, Wesnsume that S 2 ohpauhte withthe presdsnee consents Thon has the Hem (AI BD sniere J ivthe ht jo under Sand A and ae subsequences ofthe other jms (SB. ether A ow B could beeps): Consider The ne oe Since S obs the prosedence contains ant J mas he ast feasible Remember thin 9(€) monsdecteaing nC, Sine al completion ies of Jobvonic than Ihave decreasedinpasing irom 08" noon other than fa, can have increased. By consracin , chosen sich th 10) = pint are Thesfore. ven if the costo noose under Sit doesnot exceed the onto! under Thusthe makimum set ander sno greterthan under Sense me ony choote an optial schedule ofthe form required Tesspom a relatively simple matter to deuce the form the ago et, beth jabwhish by Theorem 42 mae la nan optimal sequence ‘Trasthere an apinal sequence ofthe form. (AJ) whore A sa pera lation of the other (1-1) jobs The marum soxt of hi sequence (a3 iste larger of 0), he ont completing last. an the mai ‘nur ow of completing the jobs in A. Ar cin sequence canbe found by muhing oth these terms a sal as ponies heen chosen 30 that leh the minimum forall the obs that eld he proceed lst. So eastract a opin sequence our tsk m Smply to choone tha he "akin sa of completing is jobs avamalas posse, In other words se face the problem of sehedting (v1) Joby sujet to precedence enstsicts 0 that the manimam eo of he indie obs minced. ‘We tase a new probies ath the sae form as or Ong, but wah (2 Thyaheinnend of By Theorem 1.2 me can Se which ob should Be 8s fo tha new problem and, enee, (# ~ st TRE sequeece forthe ‘nial problem, eae lf nih he sb 0 che (2) ods Ani so We go on, repeatedly scheduling & Jo atthe end ofthe saguence a sa ‘Single Machioe Processing fone 0s reducing the site of the problem by 1. Eventually we completly solve the ong! problem Faample, Consider the 6/1/Eyq problem wih data Below andthe som strato that unt recede J shih nt intr peed yan alo he must precede both Je and J We show those hb 4 & ProewingTim 2 9 4 3 2 Fouling the job prvesed he ¢ = 26944392412 15, Jobe bs and cane poccsed tebe. = UB lateness wer Y= a Hence fm sched 6h the completion ume ofthe fist fs jobs y= 18 2~ 13 Jobe Jot im be proceed lst noms. V'~ 1S the minima ateness over Vint 9), (19 2) which ocais tor Ty So foo h sshd Sth ning he ob processed 4h: Wa have ao delete J and rom su lst Thus "J. becomes avaiable or processing Tat. Wo” hive eats oad VEU. Jy Mimiqum Talenese over Y= mini 99. (9'~ 7), which acc for Jy Thus 3, stale th Hinting she Jo proceed nity Be and J have now been deleted, hace we not have Wn) nd ¢ = 9~ UR. Minimum ltenest ver V= mith 6). (3) which occurs oe Dy than Bye scheded “The jabs scheduled fine am sxond ste mow automaticly and Jy respectively. feeause of the rocedence contains, These ailing ray he lak! occ usin the following table SS] USe11) (157). which wes fo hy +d, sunnot be sehedaled ios, = i possible t schedule J, ast. Sf font aeady een sh Whe hss job a ach ¢ has the minim tens (F~ dan hs falas ciied me tale he fina seeds may be fount by teading up ‘heft column on the vate of Ly, fund By aking the largest the ‘Sisk guanine Here we fi th Schade 2 DoD) ith ‘We Sal eae our tay of pcedece cons fora whe: they ate cass azn in Chapter. Now we tut tthe ope effsiony 44 SPT SEQUENCING SUBJECT TO DUE-DATE CONSTRAINTS ‘The problem that we conser inthis and the nest ston how td Sele thi are elfisent with respect to Frama Fk this set We Akoelop an algritin dv to Sm (7956) ich alough not ecto he mediate tion af the prob. mts instrrenta nthe ertrction of efficent aches. Than we al oe explosion "ficiency nga ul he pent ec; eather we fll Sith nthe ot ‘Spee tht we hive 7/1) paabein 0 se, By Theorem 3.4 ‘no thats soutin my he foun! tae EDD ele. Supe that her Comet this seed he fnd tat, ~ Wg thi al the de tes Sin te met Then Reon we actualy pris he Jobe EDD soquence phen sal sp sed think. There might he schedulesthe ante 56 Single Machine Processing [ch EDD sich als sty the de date constrains Might we peter one of these? For ntonce, oe might give #araller sae of Fitan the EDD tequence Indeed thse more than pose: interme salve an example i show this Samth'salgoraten gives usu way of finding a schecule to rine F subject the condition Tat T= 0. "Avdice Lanier Soh’ alors bud op 2 shed fom hack to feont fittings 8 oh to be mth inthe processing sequence. thea {ir Ds. sn soon. Agsin a with Lawler, iis based upon 2 theorem, Which describes the characters Job that may be proceed lat, Theoret 4.3. Fr thet job, 1 machine problem such that al he duc dates can be met thre exite schedule wich mines Fees to T= 0 nL abi fo J ast, and onl Wa Yn Proof, Consider 3 senedule Ss which ob 8 ad condos () and (6 id Spe we erg J, a ne te seg. (po contin @) Goes A old far J then Jl be tardy find «0 wil mo long beast, On the oer hand. id) =p, bat Pepe condition Ci) des no old forthe new eguenee, ten a show th theatcan flow time ofthe sequence is neease, La the “chedile with, and, interchanged be Then. wing prime we india "ana connected with Swe have ROR sin and, moreover, the os time of any oh sched hermes an J il be increased by (px Pp) = 0. The flow times ofall corer jobs will be tnehangsd 1 allo immeialy that F > F Thun the theorem i =) Tt snow sraghtfowars matter to develop Smiths Algorithm, We somit he arguments undcriving is devslopment since they ae entirely paral to howe we used when eostracting Lawler = Algorithm, Thos me ae Amit 44csanny fe oko Tene *€ ps Sak eu Pr dle Dh t Sing 2 Decrease hy VE dEEMNE # BY a dete J, fom sco 4:4) SP Sequencing Subject to Bue-Date Constraints so Sip 4 ME there ane more jobs 1 Schedle, Le. HE = 1. go 10 Sip 2 Otherwise stp withthe optimal processing seQueREe {nly tn ating the gorithm we have used the following notation 6 the postion inthe prosening sequence currently Being files (NB. eyes down (oD +s the tine at which the job Atha the sequence most compete and U ithe st of unscheduled obs, “There are owo remark hat should be made upon the working ofthis hm. Fas, inp in our development tt, hetore wing the mvt, we check the EDD sequence to se whether there is schedle thi Tay 0. In fast, me need not thin there hn schedule th Troy ite shal iconer this in working though Smith's algoitim, At ‘Sie pae through Step 2 we stall be unable to find an J tthdy. = 1 cond sometimes J may not be determined agus there my Be 0 vere jobs for which cmos) and (i) are tre. hes, me aE The chaice oJ, arbitrary at est forthe presen Fanpie Comie soing he probe ibs T= hen bb he a eet 2 9B Bac Date 5608 Applying the algonh we tke the folowing steps Sep 10 k= 44 = KU Ui do Dud ‘Sep 2: Only satisfies conto (i) 9 We OO y= Sep k= Ren eu = Uy del Step 2> and 3, satisfy condition (nas he larger processing tne 50 Joh Siep 3: £2 2, = 3: B= Uh I Sa kei Sp 2 J and Jy sais condition (has the lange prosesing time, Sed J satis onion (90 Ja ~ Step 4: Optimal sequence im (By Jy Jo Ju has F= IN. whereas the EDD sequense (14. uh) F= 4 (eopennyeu) M4 NB thesequer PeaZA tal cont = 5+ 24> 6.P = 21> total omt = 6 675021 = 171 x27 = 200 Hence the minal ot cost shed is U3 3s) Thee san itcesing and portant peomctrcinerpetaion of th colealton, which wl be amie those who have Sued thai poiemming, (Se Section Kl for rsleroners} Ip Figure 48 we have tied the amity of ines pven by 81, =F = for vying vals of Ras dscreases eve ine move aoe te igure a shown, Moscover cach fee jms pots (FP) A are cau! erm of ot cost. The pont ) oeesponding Wo seed 3) Jo) es om he ta Sst ie VRID and i ie leu that lne’of Tower otal ost poses thesgh seit sched Ts we ee mma att ede ches the mime ol cot co Single Machine Processing {ona 46 REFERENCES AND FURTHER READING Tre section on required strings of jobs followed Conway, Maxell end Mier (1987, pp. 10-71) fry choiely. Their book slo considers single toachine processing wth more genera forme of precedence oni However, for the thst general ement of n/1//F problems sith prec lence constants, nce Adolphson (1977) Lele (1973) f the iin ourte a ehe work n Section 43. Very recently Baker ea (1980) have fctonded Lavie’s approach to deal with non-ctasendy times, but their Solution doce assume that pre-empsion allowed. The problem consered by Sidney (1977) has mk in sommon witht! xed hy Lace Howe vcr the performance messire wed by Suey ispesfially relied io En, lind. and uy fal dhe nom dcereasing condition put on 9(C)Py Lanict Glesebvook (1980) has comidred singe machine pocesing une pres ‘dsce constrains men the cows ae Uscounted overtime The concept of efceney hava for mote general appiston sn opera sional escateh than justin cheng. His known Py the terms ad "li ad Pareto opimalty.A bt. bt exelent serve of te ets md their mportance ven n Keeney ang Rail (1976, Chapter 3) 8 more theoreti and deine stds pen by Geotrion (IB). Our tea iment of effwieney with respect Py and Fin singe machine problems rns Van Wateenbove and Geet (1980), Suprsngy there docs ‘sem tobe aie gemcrlfation of those owls a efceny wih report tof and 3) jaP Le. 4 eiged meen flow tenecrtoron, See Burs (1978) snl Bansal (1980). Van Wassonbone and Baer (190) mrodace furter ideas of efseny into schedling theory ad ove another las of Single mishine problems As appre! from the ates of these reerencer, Clftiency conepes heve ony Jur eared the hedulnglveretne, How ‘scr: from their api adoption in ther atexs of operational esearch we nay epee sudstanial development a the op 47 PROBLEMS, 1, Solve theftloning 12/17 problem given hat obs (36,9. 12) and jot (122.48 10) mst he proceed i tings. be rasase7e9 mn Prceumg ting «2015 6 3 4.1 2 SOE the 81a prob ab se0471 Protilems 6s ih precedence constraints 6s 1478 5, Solve the 4// man (4{€9} problem: Joo pasa Precwing Tine 20 2 1 prcelence constants : nd mbewe nO=e nO-1 se res Dw tea yA) = 2omwniou 3) 44 Note tat Lawler method may be applied opens witbou ay precedence consrans In pal vee Eales git © Sow the EDD cequence snes Bo he 1//q and the 9/1/77, problem We doun Lal's Aigo fa am expla te by step Format. Uae Smith's Algorithm 40 salve the 7/1//F subject 60 Ta 3 oem wih dat Jo eb Ss ey de Prccungtim 6 28 9 3 1 8 Due Date Hoe aon BoM 7, Generate al sheds ecient with respect 10 Ta and F for the 7 ja proto in question , Hence find he mina cot shel oe AaB ee Chapter 5 Constructive Algorithms for Flow-Shops and Job-Shops 5.1 INTRODUCTION 1 is oom time to turn our atention to problems with more than one Inachine, i" scheduling in the flow shop oF general jobstop. Out Inmediste concern will be to sty problems for which constructive tdevriths exe By a constructive igri we mean one which Buiks UP SInoptimal sction tom the data ofthe problem by following 2 simple set Of ules which enaciy determine the prossing order. In the singe Imashine case we encountered a variety of problems which could he solved Constructive, We shall ot be vo fortanate here; only very few with (80 ‘Srmore machines are amenable to such analysis: Indeed, the %/3(C)F me family of problems isthe only one Yor which there enss a constractne grim splicabet all cases. Fo al other fms the few constuctne “gorithms that xi apply only to special canes usually eases in which he processing times ae rested in some way. Im hiehapter me consider the host important of tes gorithms, namely thowe duet Jobason (1954). "We make ony one sumption over and above those made in Chapter 1. Wesasume thal jobs ate walle for proceosng me datsy.This Assumption for Chapter § Grn O fora yd 12. oom ‘82 SOME RESULTS ON SCHEDULES IN A FLOW-SHOP In this section we consider two very general seals about the form of Scheduler in'4 Bowshop. Int lwahop the technologial constraints ‘demand that the jobs pas betwecn the machin in the same order ie iJ, ‘must be procsed on machine M, before machine My then the same iste fora jobs, We shall ee that these general results imply that for some 2 fn'3 machine fow-shops we ned only consider permutation schedules. A ‘Sec. 5. ‘Results on Schedules in > Flow-Shop_ 7 permutation schedule is one on which each machine process the jbs in these oder ion machine Mj, 8 proce eee he a lowhops thee is naturel ordering of he machines aml that given bythe tchnolopeal constant the processing onder foreach and {Ese 08 Thos we shall sume that each jb must be prosewed DSM, besote M, before M, ete The fechologal constants, therefore, hae he foun “Table me technoogial cone ‘n'a fmeep Joh Procening order yO MMMM 4 MoM Mio Me 4, MyM. Myo My ‘Theorem S.1 Forth nm problem with B regular measure of pet= fovmance, ie sufcent to conser schedules in which the same rocesng saquence ie gven on the fist two machines. Proof Ifa schedule does not have the same order on both machines M, {and M, there isa ob / which ctety precedes job K on M, but follows pethags with intervening Jobs, om Mb. (See Fig, 5.(@).) On M, we may fever the order of and K witout istetsing any stating tne on M. Thar thintrehange cannot increase he completion tie ofa job a hence, neither can increase a teglar measure “This proces exchanging jobs oo M, may be repeated uta stele Ss obtaines with the same order on M, as on M, Clearly Sf no Woe than under regular measure ‘Theorem £2. For the nim/FFau problem, there s no need 10 conser schedules with ifferent processing orders on machines Mand M,, Proof, I schedule S, does not have the same order on both machines Moy and Mc. there ij 7 which del precedes joo Kom My, but foloms k, peta with some intervening jobs, on My, (See Fig £108) Sipps we severe the processing ander of and K ta M,- Clearly thi nay change the flow ties of india jobs: some may increase: some 68 Constructive Algorithms for Flow-Shops and Job-Shops[Ch. 5 Sri ie seer " ii ee i. tLe (@ Mesa, may decease. Homever, in total the processing on M can only be ex edited, because J completes on M,_, before K-08. THUS Fg NOt ‘increased bythe itrshange. ‘Ths process of interchanging jobs on Ma, may be repeated until a schedule Sr obtained withthe same order on My as om Ma The above ‘shows that 8 eam be no worse than Sy terms of Pay The proofs of Theorems 5.1 and 5.2 are very sma, but by no means the same. The sinilanty evident from Fig 3(a) and (by, these der ‘only inthe labeling ofthe machines. Te important dinction between the proofs es inthe machin chosen forthe interchange of processing ‘order: In the proof Theorem 5.1 me reversed the ender of ana Kon Mp the frst of the two machines considered, We saw that doing so could Improve and certaily ot worsen each and every flow time. Ideally we shold have liked to copy this intrchange inthe proof of Theorem 8.2, fo have interchanged Jobs on Mc. For then that theorem would hive been aplicabe to problems of sehsdaling aginst any regular measie and not jst Fa Homevera ite thought shows that the effect of trehang- Ing obs of, is entiely unpreditable what happen depends wpon the times jobs complete oa M, Thus we mast make the interchange on Ma and so can only deduce a esl in terms of Pre ‘Because of Theorem 5.1 we need only conse permutation schedules for the /2/F/8 problem, when Bs rear. AdGing to this the result of Sec. 5.3] Johnson's Algorithm forthe n/2/F/F gq, Problem o ‘Theorem $2, again we need only consider permutation schedules for the AISIF/F problem, Examples given in Problems 88.1 show that these ‘result cinot be stengthened 10 more machines, |5.3,JOHNSON'S ALGORITHM FOR THE 1/2/F/F =, PROBLEM ‘Suppose we are faced wih ann)» problems: we have to proces obs throug wo machines. each fh Inthe order MyM sorta the ‘nanan flow une i minimized Not Ua, Secs i ole have 70 Now itscms sense 10 tart he processing with the job wih the shortest processing time on MF then the processing om 1M, may stats soon as pouuble (See Fig. 5.2) Silay seome jot a6 sedsile o fish the processing with he Jo thats the shortest process ing neo M, For wll dij ik proceaing M, mit be els we ave just shown tht we only need to consider permutation schedules for this problem. Putting these Meas together, i seme reasonable to sgest that he optinal schedule permutation of {1-1 sich thal the ‘lor jobs in the processing Sequence have short Mi, roceming ines, ‘ete the later jobs have short M, processing times. "The schedule Com ‘Sete by Johasa’s Algorithm has precely this propery 'We shall give Johaso’s Algorithm next and an example of its we ‘Ten we sal prove tat his method does find an optima sched 1 i simply our notation tole 1, Py the procesing time ofthe 3,00 Mi: and b= po the procesing ne ofthe Jon M. The algorithm builds up the procesing sequence by working in fom the nds towards the middle. To do thie we shal eed two counters, amely& My matt be le whe the ft b's process on My fli sees] sce ESET is lost eb 6 proce " (eee ne Rg 52 The ie ie at tf ay chee 70 Constructive Algorithms for Flow-Shops and Job-Shops [Ch 5 and 1 A = 1 inialy and increases 2.3.4. a the Ist, 2, 3d 4h ppsion in the processing sequence ae fled. Simary. = inaly and Aecreaes (nT). (r,s as het (0 1h, fw 2Mhs pone tins inthe procesing sequence ee filled Abert 5.3 ehmson) elu °]/F/ Fame Sup 1 Seth = 1,1 ‘Sip 2 Set current In of unscheduled jobs = (Je) ‘Sip 3 Find the smallest ofall thea a0 , tines forthe jobs curently ‘omehedes ‘Sup 4 I the smallest time is for 00 fst machine basalt, then 1) Schedule J tn kth postion of processing sequence {i) Delete J fom current is of unscbedulea jobs (i) Increment to k +1 (W) Goro sup. ‘Sep 5 Ite sles mei or Jon second machine, Leb smallest, then 10) Schedule Jn the fh poston of processing sequence. (i) Delete, fom curent eof unscheduled jobs (Gi) Reduce (10 01) (i) Got supe Sep 6 I there at any jobs sill unheduled, go to Sep 3. Otherwise, stop alles ime oc for mote than one jo in Step 3. then pick arbitra. ample, Schedule the 7/2/F/Fu problem with dt Processing Time on Machine Jeb My My 1 6 3 3 4 3 s 7 1 6 4 s Applying the algorithm the schedule bd up a follows Sec 5.3] ohmaon’s Algorithon forthe /2/F/ nu, Prublem n Job 4 scheduled: 4 Job scheduled: Job 2 scheduled: 4 Job 3 sehedoled: 4 4 4 Job 1 schedule Sob 6 Schedles: Job 7 schedule: ag: ra o71 “Thay we should sequence the jobs inthe onder (4,2. 67.1.3, 5). 'NB. Inthe above there ate two arbitrary ches, We coal hove put Jol 5 into the last postion ofthe sequence before scheduling Jab Here ‘he resting sequence would have heen the same. Also we sould have scheduled Jo Tin the sith positon instead ofl 3 This woul have tea ‘oadiferet butequivlenprocening sequence, is (4,2.6,7,3. 1.3) ‘We now prove that Johmn's Algorithm does ree an opie schedule We begin by showing that the ist cle of algorithm chooses: tnd posions aj opin ‘Theorem 5.4 For the m/2/FF problem with, Denn and = b= (0) Wa, = mings, as --- t BeBo vss Bas thee 6 an optimal Schedule with fst inthe processing segucnce (0) Mb, = mifay ayes 0 Be Bees. als there isan optimal sete with, lst the proesing sacnse Proof We prove (i fist. Let be such that a, = mina. tr by barb. Let bea chee with tft inthe rocesing sequence LEU), e te job which proceed immedstly tore J Then 9 has Gant ciogram as piven in Fig. 53. Let Q be th tne thai M veal start processing J under 8, Let Rb the tne that My finishes proce {ne jo Before J under Sand is thus teady to process J ‘We consider the schedule $ obtained simpy by interchanging the bs 4gand 9S. Our aim sto show hat, = Civ nhereC, the completion Hime of, under and Cs the completion tine of under SH thst Fn" Con ™ Con Fe and We deduce tat (I) Rois by repetey ineschangng J, ith the jos Beto tun ts fst fo be paced Unit $3 starts on M, at max{R,Q + a. We know, by assumption, tht a= by Ths completes om M alter, completes in M, and 0 we G rman(R, Q +4) ++ Unies SJ, stars om M, at max(R, Q +o) and starts om M;either as Soon as, fines ora soon as M, thes processing J, whichever he n alo “Ifa sooner. $0 C= mania R.0 + 04) +b O05 +a) + be Hence C)= maximax(R.0 +04) +6, 0 +a,+e+b) (1) Now C= maxlRQ +a) +04 by E minlRLQ + ai) +B, + by, Because a= 4, (by assumption). = maniR.Q +4} + b+ By Borathsd, S Ora rd +6, because by =, (by assumption). SC. iso les than either ofthe terms in the water mas of (5.1. Ths C= Cas required The proot of (i ef for Problem 5.82. snd tis pow a eatvely simple matter to deduce that Johan's Algorithm also positions the remaining (» 1) jobs eee ‘See. 54] Johnson's Algorithen for the 9/2/G Fu, Problem B ‘Tocorem 5.5. Algorithm 5.3 finds an opti schedule or the 9/2/F/Poa problem Proof. By Theorem 5.4 we know that the fist cycle of the algorithm poutons 2 job optinally” We shall show that if the fit» cycles have Pastioned » jbe optimally, the (+ It eqee wil abo poston job pinay. Our esl then fellows by induction Suppose » eels have positioned » jobs optimally. Because of this ‘optimaty the ("+ Ist eye ned nt reconsider any ofthe choices made eae cycles ay spy ehoose one ofthe remaining (r ~») 4008 ‘nd poston sin ope ofthe remaining (n+) places ia the procesng “quence. What the cycle actually does isto choose and postion job tecording 10 the conditions of Theorem 5:4 applied tothe remaining (qv) scheduled obs aloe. Look span atthe proof ofthat thoorem ‘The argument shows that C, > Cor the interchanged jobs. Moreover, oni anus dicey felting to and) enter the argument; the rest of the pocesing sequence Hrelevan. It follows that ia placing’ the (Crs st job to minimise Fa, forthe remsining(n ~1) jos cosiered Bone, the algorithm is minimising Fy. fr all obs. Thus the (e+ st ‘ele postions a Job optimal We know tha he first ycle postions a job optimally Setting = he above shows thatthe Second eye also postions 3 job optimally. Next, Suing = 2 we deduce that the Ud jl does so 00. Continuing ia hs ‘ray deduce that alleles poston jobs optimally. Tus the algorithm ‘commute an optimal schedule 5.4 JOHNSON'S ALGORITHM FOR THE #/2/G/F.., PROBLEM, 4 Section 1.4 we made the assumption that each job must he processed "rough al the machines. For thi setion we drop that esumpton. ‘Suppose tha the st of m jos (J.J J.) may be partion into four types of job a5 flows “Type A: Those tobe procested on machine M, only. ‘Type B Those tobe proceed on machine M, ony ‘TypeC: Those to be processed on bath machines the order M then Ms “Type D: Those tobe proceed omboth machine inthe orer M hen M, “Then ite thought shows that he constraction of a optimal schedule seaighorward (1) Schudule te jobs of type Ain any order to give the sequence Ss. (2) Schedule the Jobs of type B in amy onder to ge the sequence Sy (3) Schedule the Jobs of type C secording to Johasons Algrith for IBF problems © We the Sequence Se. 74 Constructive Algorithms for Flow-Shops and Job-Shops (Ch. 5 (A) Schedule the obs of type D according 10 Jonason's Algorithm for IRF Fn pobims to pve the sequence S-(Bemare: Here Misthe frst machine and M, the second i the technological constrains) ‘An optimal schedule s then: Machine Processing Order M, (oS) M, (580) “To see that this schedule is optimal remember that time 5 wasted and hence Fa, sinceased, either M's kept le walig for jobs of type C19 ‘complet oe Mor Mis kept idle waiting for obsof type D to complete on My This schedule clearly minimises sch idle time, Bxample. Consiter the 9)2/G/F au problem with times and procesing ‘onde given by Proceng order and es Job Firat Machine Second Machine : M8 M2 pees Mace MS 5 om 8 Ma qe Meee M7 SEEeee Meee. M4 aaa M3 7M 8. i see = a oM s - ‘To find an optimal schedule “Type A jb ony job 7 st be processed on M; aloe ‘Type B jobs jobs Band 9 require Malone Select abtary ores (8 9) ‘Type C jefe 12,3 and require Mist and then My Johnson's ‘Algorid for this 42)F/F problem gives the sequence (32.0, ‘Type D jobs jobs, 6 requte My ist and then M, Johnson's Algorithm for this 2/2/F/Fnq problem gives the sequence (S, 5) (Remember M, Wow the second machine) sec 55] A Special Case of the /3/F/F 4», Problem 1s % A] b a sb Ll imal seauence forthe overall problem is Processing Sequence MochineM, 3.21 75.6) Machine, (5.6.8.9, 4.3.2, 1) andthe sulting Gant dagram sgn a Fig 5. From this we se that Foy 44 foram optimal shed ‘55 A SPECIAL CASE OF THE 9/VF [Pq PROBLEM Johnson's Algorithm fr he n/2/FF aq problem may be extended 10 a Special ese ofthe m/3/FF., problem” We need the condo tha: iter sing mie) (2) « a= ed ‘the marimam processing tne onthe second machine 0 grate than ‘he minimum time on ee heft the third 1F(5.2) holds an opal schedule forthe problem may be found by leting b= pat Pm and scheduling the jobs a8 they are tobe processed on two machines ‘only, But with the peocessng tie ofeach jo being and, om the is land second machines respectively Remember we have shows that we eed only conser permattion 76 Constructive Algorithms for Flow-Shops and Job-Sheps (Ch. 5. schedules forthe n/3// aq problem and note that Johnson's Algor Spplied tothe n2)F/F problem withthe constructed a and ties dows ‘produce such a ached Tak you to prove ia Problem 588 that i, Indeed, an optimal schedule. In Section 73 we encounter an example, stich shows that condition (52) ners fr ths method oid a ‘Optimal shed Example. Consier the 6/3/Fa prob with data Fist we check that (5.2) holds for this probe. Here we have font =3 ant i = Fe SS Gut gran forte FF te ‘Sec. 5.6] Akers’ Graphical Solution to the 2/m/F/F Problem = 77 “thus we have mit {py} = 33 = max {pa} and (52) holds. 8. (2) mancitheror condition: we do ot need oth inequalities 1 hold.) "The contruted and D, times ate pven in the be. Appiving [Algorithen 83 es the sequence (2 4, 8.1 3, 6), which has the Gant Gingram pon 0 Fig SS 136 AKERS’ GRAPHICAL SOLUTION 10 THE 2/m/F/P,., PROBLEM. ‘We have now examined ll th consutve methods for prodcing optimal ‘Shel that we shall ia this book. The lgoritms im the remaining Choptes are enumerate oaes. They generate schedules one by one ‘eatching fr am optimal solution. Often as we sal se. they wseelever ‘mination procedures tosce the non-optimaliyof one sede pis ‘he nom-ptimality of many eters aot yet generated. Thaw the metBods fnoy not search al the sot of feesble‘olutions, Nonetheles they are Incthods that proceed by exhaustive enumeration and, hence, require ‘hack computation. Ava fist spin thi move tthe stad ct numeration tncthods, we consider egrphical method doe to Akers (1956) ‘We somider a2 jb, 6 machine flow-shop problem 2/6)F/F yx The case for guneral m machine flow shops follows acooraingy. In Fig the ‘ans marked off in intervals corresponding to the prcesing times of sob? pease aa nee 78 Constructive Algorithms for Flow-Shops and Job-Shops (Ch. 5 {ob 1. Simiat the y-axis marked of in intervals corresponding those St ob 2 Noe tht the machines have satura odin ince tsa {owshop problem The rectangles ws sks corespond to the proces ing timer of jtn I and 2 on the same machine a shaded 0 These correspon (0 eridden tones meaning tht any one ace chant oes oth jobs simltancousy. ‘Achedule maybe rected thi i a by pling line that (a) runs from © to X, ie. from (0, 0) ("pp (©) iseompoted of spent running 1@) honontlly, representing work on job 1 ony, (i) verily representing work on Job 2 on, (Gi) 815, representing work on Both obs (6 doe not enter any forbidden tone, does not demand that any ‘machine process hh besitos “The cave! way to understand thi isto conser a patclar example: the schedule 5 represented by the ine OABCDEFGHIX in Fgute 55, prosessing i desrbed in Table $2. “Table $2—Desrption ofthe schedule shown in Fgure £6 SEGMENT Job 1 eo? On being prossed on M, waiting ae being prsessod om ML, —_being processed on M, ae being prcesod on M, wating cD being processed on M, —_beingprocested on My DE being processed on My waiting EF being processed on M,_—_being procesedon My FG being processed on M;——_heing prosesedon M, GH being processed on Meng proseeedon My HI being processed on M,—_beingprocesedon My rx completed eng proceed on [Notice thatthe schedule is a permutation schedule: all machines proces {he jobs inthe oder (1.2) Hew be woth yout whist stop and interpret ‘te Schedule ven bythe ine OABCDEFGHUX in Fig 87. Note that, Sis nota permutation schedule The maximum fw time of schedule i given by ether Fay = toa processing time on job 1 + total fe dang which Jobs processed whe jo 1s waiting Sp. smote over emer of edn ‘Sec. $16] Akers" Graphical Solution to the 2/m/F/F 4, Problem = 79 : i F, fon = ttl processing tine on jb 2 ‘foal ime during mich job 1 proceed white job 2 is wating = 3p, + sum of lengths of horizontal segments of schedule line, ‘Theserms pan p are independent ol he sched Thus fad samoptinal sede we ms constr the shel ie that as the east eng oferta segmcts oe cguae ay. the kes length a total ‘Spent How do we constnct hese? "nen cath schedule line represen a co or each shaded or tasden epi. may eer (a) pass above the region indicating that job 2 i procesed om that Imochine before job oF (b) pane telow the region indicating that job 1 i procesed on th Imochine Before J 2 ‘These choices are represented ia Fig 58, which shows the configure "ons which may result teach of he four eases a choe mus be mae at the poin C Should jb 2 be processed on Mf, hefore jb 1 in which case ‘he Upper branch mast he taten? Or shoul ob T be proested fast and thas the tower Branch followed? Not tha in cases (0 a (the hice obvious. In () job I can complet on M, here jo 2 request Thos, 80 Constructive Algorithims for Flow-Shops and Job-Shops (Ch. 5. ‘see. 5.71 ‘References and Further Reading 81 crenflaterin the schedule there san advantage in holding up jb 1 tet Jf 2 overtake it, there iso need wo consider that yet andthe lower branch Tpould be taken’ (See Problem $8.10) Similars in ease (B) the upper ‘ranch should be taken ‘This construction may give 2 pose schedules or, rather, 27° if we sas take ww of Theorems 5.1 and 5.2. Thus for a problem with 12 imochincs we may need to consider 2" 1024 schedule nes, However, in frst, eset (0) and (4) happen a ot, and alo the human eye i ery fro! a sping the schedule with the lest total length of horizontal ‘Sqments. Thus ts onal fail easy to find an optimal schedule simply by loking st the graph Frnally, we note an example which snicates the complexty of the schiding tsk im peneral Consider Fig 59. which shows the Ist fOr tmachnes im 2 im/ Fo problem. At point Cit seems attractive tn the ‘Bor tere to pass undo the Mf. forbidden zone. However doing #0 “rap the schedule ia a rion th ite possbty of using 45" ines Paving over the M,. forbade zone leary optimal. In other ord, when deciding the processing order of & pair of jos on 4 patulat Imichine, it necesary 10 conser the ramifications of hi choice for the thedling ofall other machines Sehedalingis generally avery dificult tas cate few nyo one's chokes ate odependent alte others |87 REFERENCES AND FURTHER READING Jonson (1954) the original source of the material presented in Sections Sies-4 hs paper hae been reprinted in Muth and Thompson (1963) ‘Thee ae many alteratve proofs of his resus: Conway, Mawel and Miler (1967) and Baker (1978) general follow Johnson's approach; Ri roy Kan (1976) ives entirely dllerent pons, upon which ous ae ted and White (1968) uses dynamic programming dens wo derive the same ress. ‘evar (1977) has examined the FF problem in deta extend- ing and enlarging upon Johnvon's resis. General condos fr optinaty hve Bech obtained for he general m machine Nw shop by Yueh Mish (G96), but his results lead po useful slgorths, Panwaller and his “oavorkers have prodved constractive algorithms for several very Special ‘cs ofthe m-ichine flow shop problem in wish the processing anes te restnted to obey many conditons on their eave magstude: (Pus walker and Khan (1976), Sruth eal (1976); Panwalker and Wool (4979, 1940, The presentation of ARers graphical method given in the previous section Ghsely follows that of Conway. Marvell and Miler (1967, p-9%-100), Because of graphial ature, stisa method coined 0180 Simensone and hence does not generalise chou 82 Constructive Algorithms for Flow-Shops and Job-Shops (Ch. 5 4 PROBLEMS 1.(2) Consider the 2/3/E/F problem with procesing times given bet. Machine ob MM, My aeseuinenl Sow tht the two permutation schedules (1) and (I) Hoth have ‘mean Now times of 15, and tha! the schedule below ha mean flow time ou10s M ra 7 M ee Ms nou Show tha both permttion schedules have Fr = 17, whereas the fehedale below hat Fon = 14 Machine ‘Sec. 5.8) Problems: 8B 3, Schedule the 8/2/F ya puobon with procesing times: Find the maximum flow tne forthe optimal schedule 4 By considering the following processing times show that Johnson's Algorithm doesnot solve the n/2/F problem oom OM, 5. Genetalse Problem 584 below in order to show that Johnsons leith may be very poor for the n/2/FF problem. Consider jobs with Drocesingtimesa,~ efort = 1.2,3,--.n:B)~ Biandb,~ delori~ 2.3, rm Assume ne < . Show fist tha Jason's Algorithm may give he roesing Saute (Jody orth seen show tat Pmn = 8 +E + (0 ~ 18/8, Next show that an optimal sequence is Jy. Jed: {Hinz Show that Fonlydependson the position of inthe sequence and ind hs dependence). Fr thin optimal sequence show Fags fin os ty OEE $4 Constructive Algorithms for Flow Shops and Job-Shops [Ch § Hence show, that 3660, FaygalFoe 2 This example du to Gale and Sani (1978), 6. Solve the 18/216), problem below and ind the opeimal mai um fo tine Processing Order and Time Jobin Machine Second Machine taeedgceaaee M tases Mos aeaaseaa ag) MD Siena M13 6 MS M8 retinas a) ates eer ctal M10 5 Mo M6 M10 Moon cea cee aca . BoM Ss a wom 7 Sree 7 Solve the 6/3/F/F ua problem and draw Gant diagram find the opsimal Fon 8. Prove that Johnsons solution ofthe /3/F/Fau problem, when con- iin (52) ols, does give an optimal schedule sec. 5.8) Problems 85 Ini Let be any schedule and sume the jobs are renumbere this onder, te.8= (122.3, Let py =a Be and p= fOr deere ee Anume ins iad ~ may?) Also ats, beth staring time of}, machine j= 12,3. Show that yoney | Gabbe sot manly yt ron Hat BIE 23,000 Now feta) = 3+ fy and by = 6+ rand consider the schedule S forthe FF problem with these constructed processing times. Let, be the Sorting ie of Jom machin for his 2 machine problem under S sown tts DB andar 4 wie Hence deduce that maximum flow time under for orignal problem = ‘maim flow te under for constructed probe ~ 6) Since the {hom Scrfusindependen ot, minimising the maximum fom time forthe ‘Smiruted pro sao minis i forthe orginal The proot fr the {ae where min fy} > max 8) is sir Explain carfuly the interpretation ofthe echedle Sn Fig. 5.7 Why would one ot need to consider hit echedule i ooking fae a8 opt smal chee? 10" Show, by considering Fig 5 10that incase (6) of Fig. 5. 8the lower path may always be chosen at the point C 86 Constructive Algorithms for Flow-Shops and Job-Shops 1, Sole the 2/6/F/Fa problem with processing times: Processing Ties Machine Job Jeb M * 3 M, 6 7 M 2 4 M. 5 2 M 3 5 M, 4 6 fom Chapter 6 Dynamic Programming Approaches {61 INTRODUCTION [As temarked nthe lat chapter, we have now completed our dscusion of structive methods fr solving scheduling problems and we pass on t otsider enumeration methods These simply Is, oF enumerate al pos Ibe scheduler an then eliminate the non-optimal schedules from thelist leaving those which are optima Inti chapter we dues one such Ee ‘rl approach called dynamic programming ‘Dynami programming orgasm Belinan (1987) and s appli: abe very many opumsung problems, not jas toe sing inschesling ‘White (1969) provtdos ap introductery retest and am extensive review ‘fis applications ako mest general Operational research textbooks i fuss ely. Rowghly, the method apps to any problem which cane broken down into asequence of need prblemy, the slain of one beng ‘derived na stsgh-forward fashion fom hat ofthe preceding problem However we shall aot dim the dynamic programming approch in 30) feat sencraliy. but rather look at specie application ithe Feld of SSteduling, and allow. I hope, our intuition to brat the underying ‘dew, 62 THE APPROACH OF HELD AND KARP 11 1962 Held and Karp followed the lead of Bellman and applied dynamic Programming eas to sequencing problems. Their method applies single Machine problems whore the peslormance mente takes the form EPyn{ Cy Here the {Care somes tobe nondecreasing functions of ie completion mcs It follows tmestly that this peronmance meow te is regula At purl instances we note that C.F and Take this form (Set y4€;) equal © Cj, (Cr), apd max{C,~ dO respec: ively) Tha me may ake Held ane Karps method to sive m//C./ 17. nn 91) probes. OF eoure st would be fol to eso the fst ‘wo problems; we know that an SPI schedule soles those. Bul we have 6 u. Constevetive Algorithms for Flow.Shops and Job-Shops. Solve the 2/677 problem with processing ines: Processing Times Machine ob) Job M + 3 M ‘ 7 M 2 4 M, 5 2 M 3 5 M, 5 6 {cn crapteré Dynamic Programming Approaches {64 INTRODUCTION {As rma the last chapter me have sow completed our dscasion of “onsractve methods for solving scheduling problers snd we pase om 8 “onsider enumeration methods These simpy lst or enumerate, all pos ie seheduler and then eliminate the non-optina schedules rom the ht leaving those which are optimal. In hs chaper We discuss one such gen ral approach calles dynamic programming. ‘Dynamic programming onginats from Bellman (1957) and is aplic- bie 1 very many optimising problems, not jt thon aren in shed. ‘White (1969) provides a mrosuctry treatment and an extensive rv ffs applicators also mont general operational Tesctch textbooks de fuss ity. Roughly he method applies o any problem which can be Troken down into a sequence f ested probiems the solution ofa Being Cerived i a staightforward ation from that ofthe preceding robles, However we sal mot inca the dynam programming approach in any treat generality, hut eather look at 2 specie aplication m the eld of SSheduling,and‘allow, T hope, ur intiton to absiract the underlying ica, (62 THE APPROACH OF HELD AND KARP 1m 1962 Held and Ka followed the lea of Bellman and applied dynamic ‘rogramming ideas to sequencing problems. Their method applies single ‘machine problems where the performance measure takes The forms 3 iy(C). Here the 7(C) are assumed tobe non decteasing functions of ‘he completion umes. It olows immediately tha ths performance mess lures vogular As partcular itanoes we note that C.F an T take this form: (Set y(C) equal 0 Cm (C.F and maxtC:~ dfn respec vey.) Ths we may use Heh and Karp's metho to salve /C.m/1)F. nd /1//T problems Of cours would be foo fo we ion the fst ‘wo problems; we know thi an SPY schedule solves thow. But we have 38 Dynamic Programming Approaches ben rable tose //T problems onl nw Te approach txt to sive prctlems with perormance metus of the inna 9(C) (Sse Problem 64.8) However agin dynamic progr ring t not the fest method of slo: Lanier Algor Til and Kap’ approach i based upon imple observation a the tr of optima schol Row hi ys hata ‘heute te fet joe (ay K= a aren) me form an ot ‘Shel orth redted problem basa thew K jr lone Fo {hh supp that Ons yr) a optima shee or he problem Then ee aoe ay decompor hep toms (Co Bo Gy. [Now consider the et 9s y-Jey =» Jue) Suppose that we face task of scheduling just these Jobs and nr lm, Then surely wee tmprove upon the sequence (iy Jicy--- Jun We could we would tbl to reduce the tem (61) Thus we could construct a new segue forthe fll problem ty using he improved sequence forthe fst Kobe leaving the remaining (n ~ K) nel orignal onde. The oa os of few sequence woul be te sum ofa quantity srt lee than A pl Cigna term B. But thi contacts our basi assumption that Cry “ap an pial sequence forthe fal problem, Ts ey J Sm be optimal fr the reduced problem ‘Before we can we this observation to find an optimal sequence, need iniodacevome notation, We hav already been wing he fist [St ihisinformaly. When we write (Jy...) we simply mean the ‘tbs fined between the cu parentheses there 0 implication of Putcalar processing ender However, when we ise curved bracket, Wi Gy scr J)s we do wish to imply an oder namely. J is. 0 nd tn short te forme notation refers 0 ast, the late 11 is any set of jobs containing the jo J then Q ~ (J) ithe set jobs obtained By deleting J Fom Q. Alo for the set Q we define Cg Co= Ep cc Thus Cy isthe sum ofthe processing times forall the job in. remember that by Theorem 31 we need aot consider inbred le ‘then sohing an /1//B problem with B regula, Here we are opting {pss the regular messre 50, when sebeduling us te jobsin ‘Sthe completion ime ofthe ast ob processed, whatever sequence sec. 621 ‘The Approach of Held and Karp [Next we define F(Q) to be the minimum cost obtained by sehedlng the abs inQ optimal TE contains a single jb, say Q = (3) then FO)= FA) = vind. 62) since ther is only one ay to schedule one jb and that job mast complete BC. = 7, Suppose now that Q conti K'> I jobs The, remembering {hatte lest job must complet tC and tha in any optimal sequence the fant (KI) jobs are scheduled optimal Tor aeduced problem ivoling jist those, we have FQ) ~ minir(@ ~ 1) + 1A) 6 ter words to ind the minimam cot ofschedaling Q we consider each join turn and ak how much t would cont to schedule tht jb last We Fn cur answer by taking the minima of all these posses, "Now (6.2) defines (0) forall containing single ob. Using thee valves and (63) we can find T(Q) forall @ containing js wo jobs then forall Q containing jst three jobs and 3 om antl me eventually nd (Js) doings we also find an optimal schedule Perhaps it Is easiest 1o Se all hs in the context ofa particular example sample. Solve the 6/1//P problem wih dat: sh EAN Processing time.p, 8 <6 107 Due dates 49 16 16 Here 9(C)) = L max{C,~ 40) Fin, we calculate F(Q) forthe four sets which only havea single snemer For instance by (62) we have P(N = 1(C,) = £ max(0.C, ~d) = | max. 2, i) “Thus we generate “Table 6.10) for he four sage job sets e eee) pnd 6-3-6 9 T@ 0 0 0 0 Nex we eaeuate F(Q) forthe si sets eintaning just wo jobs. For 0 Dynamic Programming Approaches Instance, if Q = Ji} we have Cy =p, + ps = 1 and by (63) F(Q) = mingr (+ 1.0) + vl), i Soin 0 + 5 Oo The clelations fr the five ote two-member subsets are id ‘Table 62 Table 6.2—The calealaton f T1Q) foreach twee set e (3) Dodd Gnd Yad) Oo Oy 4 ts jobin sequence Co FQ Hy) + fC Minion 10) se 62) ‘The Approach of Held and Karp a” ‘Table 6.3—The calculation of F) for each threo set iB (ASH od Dedede) odode & 2 21 2s 2 stat jobinsequenceh I LAL LAN NS ce) ea dag ere ey re-Pieney 2P2 Fah sR RET Minirwn : tee ro) 2 2 ' i Finally we caleuate F(Q) forthe set containing al four jobs, wi, Q= [Jody y) Here we have Cy = p, +p. py + pe~ 31, We cae Ine (0) in Table 6. The consructon of hs exactly pares the 0 prevons bles. “Table 6.4—Thecalelation of 1Q) for the etre st of fou “Thote are two poses for each Q corresponding 10 which ob scheduled lathe columns ae divided accordingly. We mark the co ‘Sonesponng othe mimurs (0.3) by a asterisk, choosing arte ‘whenever nesesay ‘We now csleunte P(Q) for the four sets comsining tee jobs instance iQ = (Jo J me have Cy = p+ Ps py = 24 andy ( inf 3) + 24). 39) * 124), Fe b)) = 28) in(0 + 2, 1/2 + 15/8, 0+ 1018) A “The cleulatins forthe other thres-member ses ae Ii ou in Table The construction ofthis exactly parallels that of Table 6.2 except fr Qe are now thre possible laut jobs an ws each olmn abide e (nJn Jed Gq 3 Sy last jobinsequence 3, 3 3s Jy nlCo) eauy T@-V+WCd 6 Fs F Minimum : ro) 5 Thus we have found that scheduling the four jobs optimally ghes Ilaimum mean trdnes of 5. Moreover tan exy mater to find the ‘ptinal schedule. Te esters i the third colon of the Table tel ws thatthe minimum mean tardiase i obtained when J, i Schedled lax ‘This means that (J Ju} are the as thee jobs prcesed. So we look st ‘Table 6.3 and find tat hese shoul be schedule last. Thus we know hat 1) are the frst wo jobs procesed. Looking at Table 6.2 we find 9 ‘Dynamic Programming Approaches ( that J, shouldbe processed ast, This leaves J, 8 the Jb int proceed Sr optamal schedule ya J) This example as ilustrated the basi principles of dynamic pro ming, but here ae some sublets thats far Rave gone unremar ed The justification of the dynamic programing approach depended ‘our boing able decompone the performance measure asin (61), pted here Sew w= SrakCud + S rah = A+B Gay). We noted that in any optimal seguenee Bay J, Sequene Ose Jes Jg) ube opti Oe St bs Trane nui be pose to reduce A and enc the ave ‘owt lok thi argument more clay, Impictin ts the sum that rechedng the Js Uppy vos yn ammo alec the B= Spr) When we conic cine tmetaling and ready tics th ye However tender obs Ory Je soya they caplet at tine Spy Thus he completion ines Cy fe inthe tial sues dig) Ian =n, Ju) a independent of seauene we we for ob iy Jayna we an a im A independent of er Bi iis Inependnceeceron fated the sym programming apeouch fa To emp thi Problem 865 In tht poem Held tnd Kap lpr moa fm atempt to allow for aomzto fea Hines But the nracton tonze ready tines esroys the indepednce of Aan Band a ample shows thatthe mafed metho may al ond the Op sotton Look spa athe decomposition (6.1). From the independence dion we know that sn opin schedule mimes the term A Hees {ist kos are ao the rat sequence ose the reduced prob ‘sed upon thse & ob aloe, Eguly we could arr tht an ot schedule mus mnie the tem B.S0 we may abo dec hatte 4) joosm an ota schedule re do te ih sequence 0 the resce potem ted pon ese (n~ K) Job one, Bu bee tmstte cae Theterm B smmimbed sj othe onion at {teas (n~ K) obey sar before 3 py te: efor all he Jobs have she. Tha tbe preci, we sou) that as (n= Jobs also solve the reduced problem subject to ths extra cont Novetcies wea) te the oberon 1 develop an rat ‘Symamic programming sation othe Sy, roblem ‘Sco 6.3] Computational Aspects of Dynamic Programming 93 For any et of jobs Q define 50" Spr Ths ith aris that any join Qa stat al she jobs non Q as be processed ist Next deine A(Q) tobe the minimum cst incurred fom ‘Sheling the jobs in subject othe onion that ane may start bone 15 Nove that tere is no comtbution to () fom the jabs ot in. or thy single sets we have (0) = a(1)) = hem) +P ee swon=>(0) wo (Generally we find 8(Q) by considering each job in and caleulating the Cont of sequencing ths Job fs. a(Q) the mieimum ofthese costs vie 810) ~minla(O ~ 1) + rhe 9D 63) Noxing that Usiog (64) and (6.5) we may build up the values of (0) fr al sexs @ Fist (64) gives a(Q) for al inglejeb sets: then using (6.8) we obtao {A(Q) for all two-jb ses then forall eeob seis. and 50 on. The procedure isentiely analogous othe manne it which webu up T(Q). For an 1/113, problem iis more natural perhaps, to base Ou she ‘ion upon expeesons (62) and (6 3) for T(Q). Doing 30 means tht wrk forward through schedule forthe fall problem considering subse uences of jobs pruned fist. For this reason oar proach # called forward dynamic programming. However, for oterprebems i often ‘more natural to base out solution on expressions analogous to (64) snd (6.3) and conser subvaquences of jobs procested Ia. Such ab proach ‘scaled backward dynamic programming 6.3 COMPUTATIONAL ASPECTS OF DYNAMIC PROGRAMMING 11s apparent that ou dynamic programming solution wo the 4/1//T exam le ofthe lst section required Too elation far more than any ther Tolution method that we have met. Admitedy” each sep lavlves only Simple addons, subuactns, and comparisons: bul the numberof them ‘oct pose the question: eit all worth. Perkape t would have been Simpler and quiher wo have wed complete enumeration, eo have lite tery pomible schedule, then calculated P foreach, and lay picked one oe ‘Dynamic Programming Approaches t ‘with the smallest T.So fet us consider the computational merits of. plete enumeration on the one hand and dynamic programming on ther. ‘We shall count the numberof mathematical operations that a com "would have to perform in order to solve an/1//T problem by each [By a mathematical operation i meant the addition of to numbers subtraction ora comparison. We shall not ned multiplications o dv but they oo are mathematical operations. Very roughly the time a ter takes to solve a problem i dirty proportional to the number ‘mathematical operations that it must perform. (Compute scientists pl forgive our naivety here.) Thus the Better method of complete enu tom and dyamic programming isthe one tht involves the lesser nu ‘of such operations ‘Suppone that we try 1 solve an /1//T problem by complete e tion, Thete are n! posable permutations of (Jj. Js ---. J.) and hence posible schedules For each ofthese we must caloulate Tor, rather Since to find a minimum of che se of mean tardiness there is no need “die by the common factor n “To calculate the total tardiness a7 for any schedule we must wie algorithm something ike the following: Algorithm 6.1 Sip 1. Set =0,Cuy)= 0, E= 0. Sep 2. ncrement k by 1 SHEDS Cay = Cn Pan Lamy = Co Sip & AY Tat © Suep 5. Ask = n? 130, stop. If no, go 0 Sip 2 ‘Sec 63] Computational Aspects of Dynannic Programming 95, ‘To find the minimum toa ardines we must use an aot some- thing ike the one below ssune that he et of alln! pose schedule lndesed by. Algorithm 6.2 Sep 1 Set Za” Bae ‘Sep2. Inegmeny by 1 Sep 3. San EPI 0, gow Sep 5 Sep & Reset Sa to 3, and vset 10 ‘Sep SIs) = (i)? 59, sop. 1 no, go to Srp 2 Ste an 4 cose hat savy ql the sales tot adie cunredso fat Atthecal of hag the ines opi Sede and Eth nmr fal tardies tied. The ore iohes G!~ 1) adoro nrement th ines og j= {¢!~ 1) comparton 1 tee mine teint 1) compartonsw termine wheter) =n Hr) “Ths complete enumeration requires n(n + 36011) 64) ‘mathematical operations to soe ann///T problem ne Nxicomer the dam programming Spproach, Here me salle (0) ea = aur of sitint subsets of exactly K jobs whieh maybe formed ftom the total st of mo, hg: Tay = mae) For the sche Oy Hae =) thialgoriti cytes through hej In heir processing Ode, or each jb i eats the completion the lateness andthe arden, finally adding tho tardiness nto the rl tora Thus tthe completion ofthe eye i the tot adines of ‘Sheole leery ths sport involves the folowing operations we say that x subst has ie K whem it contains jobs then we ee that ymamic programming calelates FQ) for dations to increment the index & through & = 1.2 ‘Stone to caleulte the completion times. bration ealulate the lateness. ‘Somparons to aut the tardiness fddtions to acumelate the rling tral ‘n_somparzon to determine whether =n & (j) sbsrs@ oie, (0) stsets0 sie 2, () mspen tse, “Thus 6n operation are rei eluate the fora tries fo ec ‘hen! pombe schedules, on (() s9bsts0 of sien 96 Dynamic Programming Approaches: (ch. We shall ot explicitly write out the algorithms needed to perform ‘alculatons foreach subnet. Intend we note that to contol k “round which the calculation cycles N times regres 28 operations, inerement and {comparison fr each cyl. (Check this inthe algorithm owe) Now for any ribet Q of size K me ft elute Cy, then me eh [Rimes comiering in tar he possibilty of scheduling ea ob nO {(Lank back tothe example of he last secon and, in parla, the Strcton ofthe ables) Thus the following mthematieal operations be performed 1 akition to find Cy (NB. ony 1 addition sneded se the example of Section 3.4) 1 sbractions to find the latenes of cach possble lst job, Ln Cyd 1k Sompattons to find the tardiness of each posible last Ta axil, 0} ations to frm each sum: F(Q = 11)) + (Ce) K comparisons to ind the minimum ofthese sums F(Q) 2K operation to eontol the lop which considers Unt the Dty ofeach jo in 2 being ehedled as eT Apart ftom counting these operations, we nee remember that he ‘sanather loop which takes us from one subset Q to the nex. starting Subst of size 1 ying through al thee, then moving to subsets ot ‘two, and o on Tiscan be dane by a oop which involves 2 operations heh subwet Q (eee eg. Baker and Schrage, 1978). Thus the dye programming method requires (6K + 3) operations persubset Oo ae fn total teretore, requires (hor sas (Yfe2+ + (3foa airs (ores) ‘operations. Itcan be shown (ee Problem 668 hat thi sum en" +3 0. 3 ‘We are now in poston 1 compote the number of operations required Complete enumerstion ss opponed to the number required by dt programing. Table 6S gives the values of expressions (66) and (6.7) ‘Mos! people ind the numbers in Table 6.5 gute starting when sei them forte first time I's computer takes T meresecond to perform ‘operation (microsecond ~ 10 second). then complete enumeration take ten milion years to soe 2 20/1/7 problem, whereas dynam gramming nll se the problem in about T minute, Wh) & dynamic ‘sec.6.3] Computational Aspects of Dynamic Programming 97 ‘Table 6:5—Comparian of the computation required 1 slve ‘ann/1/T problem by compete enumeration aby dyoamic eosramming ‘Number of operations required by Complete eoumeration Dynami Programming . 237 12286 x 108 53789) 202.902» 108 6396 «107 40 1e3 310% 1382 x10" Essentily complete enumeration fist ists all he possible schedules sac oaly when it hes done this doos eliminate any. Dynamic program ‘ing eliminates many possible schedles at onsracs te lis. Thus in the example ofthe last section the method bad eliminated by the second Suge (Table 62) all processing sequences beginning. --)- Ws ==), Gadi) ~) Oy ~s Onda») and Oy Jy ~,~). These sequence covrespond othe unastersted columns nthe abe Since there tre two pose sequences with each ofthese begining, dynam pro: {Pamming has eliminated 12 non-optimalprocesing sequences at this ary ‘Sage ofthe clelation. There are only 24 pose procesing sequences in ‘he problem, othe method i guning considerably overcomplete enume. That dynamic programming iso be prefered to complete enumeration ssa method of solving these problems clear Ie aver. Moreover, the Inge the problem the more dramatic isthe ain ia speed. However before we pt too overjoyed at finding @ method of solving ur problems, et a= Sop and think. To solve a 2/1//7 problem dynamic programming may tuke about a minut, iwe assume the computational speedo | operation Per microsecond: but 40/1//T problem wil uke over 4 years arto eng for most computer Budgets. So dynamic programming only gives sa way of saving problems with upto about 25 obs, There sa furthermore df problem associated wit his dynamic ‘rogrammiag method. It has o store and remember very many intermedi te catealtions. For stance has remember foreach ubet Q which | to sehedule lst; e. i has to remember where the asters ae inthe {abies Fr only when all he F(Q) have been calculated, it pose to step hack though the ables picking out the appropriate atensks and 0 finding an optimal schedule. Thee are 2" awensks to be remembered (iy?) Computers have finite memories. (They can only remember era numberof things) For sal computes the it isabout 2" and for 8 Dynamic Programming Approuches [on. acgr ones may stretch vp 02 So this putea farther limi on he ie probe that dymamic programming can Be tse to solve. (NB. Lam making an implicit Ssincton here between coe” tr snd'magnti de torage, Ita magnetic Jer wed, a computer can ‘ery many more quantities thn 2" However, reading them back rom [every sw compared with ecaling therm rom core. Thus using mage ‘ins help sce the problem of slorge, but ado the problem of te Inoonelsion, we sce that the dynamic programming approach nab 10 solve small n///37, problems, but it des not offer us 3 way tacking larger ones, However, pethap ll fot Ton. Consider what pens hon we inte precedence connaints {64 DYNAMIC PROGRAMMING SUBJECT TO PRECEDENCE © STRAINTS In section 4.1L made the rather remarkable statement thatthe in tion of precedence sonstrsine into» probe, a from complica ters can actualy make Rear to save Let me now show you that indeed 0. (Consider 8/1//P problem with data: Joo Hevea Procesingting,, 4026S Due dated, Tera ‘Assume, furthermote, that precedence constants demand that Ji pleted before either J, ois started and that J competed before J Fated, These constraints are shown in Fg. 61. Note that there i demand thats, J, immediately precedes J; oter jobs may interve o—®O sec. 6:4] Dynamic Programming Subject wo Precedence Constraints 99 ‘So let us begin to calculate F(Q) foreach possible st of jobs. Begin ring withthe single jbset, we ye imme that there no ptm onniderig the three sete (J,),J,- a becuse none of thes ean Be the fis join a schedule Thus ou int table as two Columns sea of Table 6.619) forthe 0 posible sngl-ob sete e i) ad pad 38 Fey) oO Newco temo et Cin a her ae he oi a tre state ($)=10 8 ier Honea, en ra these are compatible with the procedence constraints fo instance, 25) ‘ant be the fist two jobs many Schedule, beease , must precede J Moreover, when we conser those two Sts that are posse we find {hat we have no choice about which o Sequence Int nthe ofthe four tases; the precedence contsnts dictate the arden, Thos, tn Table 8.7 instead offen colar each subviing into two me have fou columns nly one of which subdir, Table 6.7—The culation of FQ) for the four pose woe ets 2 Wold Und) odd ed G 6 0 7 8 4 tas job in sequence bh uN QS uC.) eivaaaeoateeaee TO ~ 8) + Co) re of bo Mismum Bes Ee ro) ° i ° ° ‘As we contin we find similar redvtion in the numberof calculations that we have to perform 100 ‘Dynamic Programs “Table 64—The calculation of F1) forthe our posible threo @ Wedd Undo) odo dd nde 3, as jt in sequence alco) Foy) neo) Minin @ BodeFudd Undo Se) Und de 7 aie seers fenuewwo Jere [dh |e ‘Tas teppng bac through the tables picking out the appropiate isk. me find that the optimal schedule is Gy Jy J Jy 3). The tardies for this schedule i 13/5 "The solution ofthis 5/1 problem required po more etfor tha solution of te 4///P problem i Section 6.2. Yet. had there precedence constrains, t would have required much more Indeed, Expression (67) me find thatthe soon of job problem precedence constants requires about ro and a half ines as much Ss that of a4 ob one. So we see thatthe induction of ‘sec, 6] Dynamic Programming Subject to Precedence Constraints 101 ‘Table 6.10—Thecakcoato of FQ) forthe entre ‘set of oe jobs| e (ode Sede G 20 4, Tas job in sequence uC) FO 1) + Co) Miimu ra # nts makes our solution of n///3, problems by dymamie prog Fanning easier Exactly how much caer depends onthe partic prec ‘ence contain, but clearly substantial eduction in ficalyI pos- Ihe. Since very many practalycccrting problems involve pocedenes fons our conlsions in the previcus section may sem over. Pesimistc. Moreover, there may yet be a way make our solution of Bobs witout precedence cin ener Coe he flowing Theorem 63 In a7/1/T problem if to jobs spun id= hen tere Footie Poof. See Problem 66:10 This theorem is typical of las frst called dominance cndtns ination rer Ths tak he fom: ered wna pare then ere xs an optinal hele wich precedes Suppose thar we have an n//7T problem and sop tat We fn 2 dominance oncom apis tthe pat of fobs Ind Then we ow that an tial shee ents which J precedes Hence we may into te precedence consi that J shoud precede J, an be tre tat luton tothe moses problem ao sles the opal one. Moreover th inaction of he precedence oat rakes he problem ese 1 ‘oh hy dynam programming. Thi we examine encugh pis of bs ad fin! enough, dominance condone boing. we mig inode ‘ough precedence contin o make gute lrgepobls csly Soh *Sicby dynam propamning 14, are such hat) sts an optimal schedule in wich, 102 ‘Dynamic Programming Approaches [Empirical investigations have shown tha thi se of dominance om itions to mttdce precedence constrains nto problem ean ea to very {ast dynamic programm algorithms for solving Single machine problem (Baker and Schrage. 1998: Schrage and Baker, 1978: and Van Wap Seahove and Gelder, 1978). However thee isa diffieul. which fmt the applicability ofthe algorithm: mukes substantial demands on the ‘Sompaters memory. As we noted inthe previous section there ate many Imtermeditequantes which must be remembered throughout the cal lation, Even withthe reduction nthir umber brought bythe inode tion of prevedence sonsrants there ate sll suit tany 10 make probleme with mote than shout 25 Jobe intractable [ae] (65 REFERENCES AND FURTHER READING The material of Section 6:2 was taken from Held and Karp (1962) Lav (1968) and Baker (1974) prseat the backward dynamic programms formulation ofthe same problem. Laslr snd Moot (1959) develop Some further applications of dynamic programming o simply structured sched ing prcblems. Cory Masel and Mier (1967) and Bake (1974) dacs the solution f problems with sequense dependent =p ti, tlbo Corwin and Esopbue (1974). ‘Baer and Schrage’ (1978) paper was te fist to extend the use dynamic programming wo solve problems i which preedence relation toecur Their study maily concerned required stings of jobs of the dominance rotations to induce precedence constraints imo problem tehore a prin there ate none. This sugpeson has hee invested Schrage and Baker (1978), and Gelder and Van Wassenove (1978) domiaance condition that we quoted (Theotem 6.3) maybe found a tet some tore general condicons in Rinnooy Kan era. (1973), a Rinnony Kan (1996 pp 71-6). Both thee works warn of the danger intodcing precedence gees if too many dominance conditions Spplied smelancouly Thee vemusks and their method of avoiding tI Atel ae equally applicable to the we of sh conditions i dy propramming, Se ak Sed (7980) “Tranating the “pei and pape’ approach to dynamic programa to a compater by no meses saightorwatd. particulary When ‘Saence constrains ave volved. Baber and Schrage (197) sess ‘ray of doing thin but Lawier (1979) argues hat hee Way i tthe m ‘cent and proposes an alerting. 1 Solve the 4/1//T problem with data: sec, 6.6) Problems vos eb, bd Procesing time 9 127 1a Die date 1319 2321 2. Solve the 4/1//3y(C) problem forthe data given in Problem 66.1 were Sint) = Ofna Tr va, 5. Gonder api ami rogrrming tle the pncratntl/F prot and so deduce that n SPT shee proves an opel ston, 4 Solve te example of Secon 62 by backward dynamic prog og 5 Throughout his chapter te ave sumed tat thread ines of fobs ae er. Suppose we te to madly oo" dams repay approutho eat problems where this ot te ese WO come sage ih Us then deine TO)= FANN) =e +0) c@) = cu) +e, 16Q contains more than one job then for Jin Q tet f= male, C(O ~ ND) +P. tnd detine FO) ~ miniP1@ ~ 4) + ye): ce) wher Ji the jo that produces ‘the minimum a F() aoe, scaly this mdieaton tes o allow fo the pssbiliy hat the ast 4h J. the st @ may not be ready to sat immediatly the jobs Q {3] veined Show that nalgrin docsnr work bree ing he 3/173 /4C) example teow ($n G41. 3) 8 opti eb Procesing time $5 Ready time O41 Due date noo 4s HO) nor 104 Dynamic Programming Approaches 6, Discuss how you woul extend the dmamic programming met resented hereto) eal ‘with performance measures of the f ax. (n(C)). Derie the necesary recurence equations. Use Ye tmethed Yo solve the #1] ema problem with data Job dS Pomaingtine «5 83 6 Du dite m9 7 on Ready time oo 0 0 ‘Use Lawlers method to solve the same problem ‘Ni Lawlers method applies equally well when these ae no dence constraint In both ease you wil ofcourse, id thatthe ea due date sequence is optimal, bl you should note the relative con ‘onal ments of your two methods O sation Tt Consider ann! probes n which the processing mes are sequence independent. In particular, thn alls fr sequence depen “ep times sine these are included in the processing times Let be ‘rocesing time of immediately follows J and x, be that of iit Fist in the procesing sequence. For any st of jobs Qwhich contains job Jkt F(Q,) be the minimum time ree to process al te jos in febjec tthe condition hat the lst processed, Show that or Sige rah nd fr sets of more than one jobs FOU) = min, (7 = Uh + DDeduce that the minimum makespan forthe /1/C yu, peoblem is Hu Hence solve the 4/11/Com problem with sequenes dependent procs times given by pire raaa o 3267 ees io bea ete at 363-3 teres see. 66) Problems 10s carat EC = 0-1-6) G)G)- +6) 6 a2'=1() 20) +3) +. 4nf) Use these expressions to derive the formula (67) forthe number of ‘mathematical operation required by dynamic programming 9 Solve the 6/1//7 problem with the precedence constraints that J, ‘must be procesed before both J and Jy and that J, mat be processed [medately before J, which in turn must be processed immediately before 1 The processing ten and due dats ofthe jobs are +e aaa eats ry 6 4 82 0 3 @ 9 2 18 8 20 2 10, Prove Theorem 63, 1. Apply Theorem 63 0 the example of Secon 6:2, 50 nod smo precedence sons, Comair carefull the eduction in compos ‘ton that he allow Chapter 7 Branch and Bound Methods 7.1 INTRODUCTION Apart fiom hewrti nthods, brane and bound probaly the sion ‘connie min widely used rocheculing Like dram programing iis Sh enuneroteo teenie amy apa Tike dynamic programing. an “pocoach fo optimation which aps oa mac larger cas of problems than jst hon that ars in gr subject However we yal dscus yn the chmret scheduling! Agi (1960). and Lawler and Wood (1960) give owe gener saveys “WE sll maa the branch and hound approach ya forthe dy of the Iga structure of Syramse programming. Thins ott sy that the met isan extension of te eft frm wt Nonetheless is related fd apprecatng tht reatimsip wil beh frtber our understanding of ‘Sonate programming and serve ws antl nroducion to our say of ‘anch an Hound "aithragh our cxampls wil he Based pon fay single cinses of problem, there ate nt spec Tenitatiins to the bach and bound Irethod Thus ike mike to apcial awamptons Beyond those Ise Chapter 1 “12 DYNAMIC PROGRAMMING AND TS ELIMINATION TREE We begin by recalng the 4/17 example of Sexton 6.2 and comserng ow hruntic propemming eliminates the noncptimal schedules A ‘haste for ht pelo corresponds to an esignment fa ciferen oD ‘each of a in ping etn, Nem eae Sng” 34 ponile whl throggm the evar branehing sate Jhon in Fi 1 Inia ignore the distinction between sold and dove liner To generate ashes west watts 9 JP sequenced nd inte the by he pont or amde XCAR. Throughout. an Nin the peeing ‘Siac silt indice that oh fas fet Bw sme that position, WE eens rm aaa nn caanaana estas 5 108 Branch and Round Methods (ond ‘begin by slang 3 10 he fist usin in the sequen. Thus we ome {tom node XXXX to ane of the four mat cxcse and exhaustive posible: IXXX, 2XNA.IXXX and TURN. Newt we ag the second Job i the sequence soBraaching irom each ofthese four nok hice estes Tavs IXNX Pans fo he des JNK SAN. and FERS EXXX branche othe noes 218. Z3KA. and 24XNc and soon Pinal ‘ee again subdivide ou ponte yawning the ob to he proceed this Since there ae only foc Jo. this mediately fin HE Tas fb, ‘Thay for chample 12XX branches to the nodes 1234 and 1283, Tor teas that sould he appaten the banching ace ia Fi. 7. incalled'a tee. more paiulary cate We use ta Sons the ciming tion of non-optinal sels, a elimination tv. Comer the clculation of the mean aediness of any shod, say 2413. We tar by taking the job athedue fis, calculating cmpeon tie, stares and, bene, is contribution tT Th caeulaton san be epesented bya move from nose XXXX to mide 2XXX. Theakston ff the contribution tT of the jo schedul second may val be ‘epieseaed by 4 move on from nade 2XXX to nose 26XX. Thad and fourth we calculate the tardinewe of obs Land}. We shoul rewesent ‘ese cleuations by muves from 2XX to 241K and then fom 241% 0 2413, However aodes 242X an 2413 are amalgumated in cu diagram 90 ts move isnot Tally scated, (Drawing n the cra pode 240X ete ompliates the tigers to litle parse) Hence he clean of for any ‘ehedale maybe tepresented by moves slong ranches of the ce satin {£2000% snd traveling tthe appropriate final node. Fach tine anode encountered. a fuer contrition 17 alate, ‘With thse representation we may dc te sre of tert solution mernoas. For example. compete cnumeration sytemalialy Starts fom node 3000 teeny four times and trav t each ad ev) Tal node nding soi aut al the T= Section 6 owe ae Dasa poping again sat t XXXX; Dut ee there ae four simultaneous moves to neds INAX. XXX, 3XXX and HX, Thee ortespal whe calculation ig Table 1, which evaluate F(Q) fo the Singlemember tt, Nest consider the clsation of F(Q) fo the bee member ses. For instance, look a1 Q = (J) in Table 82. Fom our Calculation we determine that job | shouldbe sche ded last. neers of Me ‘limintinn tee. ming fon 2XXX to 21% eter than moving fe TXXX to 12XX. Thame inate by acl ine forthe former and dated forthe later Moreover me als know tat we nestor conser nes hat lie Beyond 12XX in our ree So we alo use dotted ines to Join noe XX sooth 1234 and 1243 In short. we hae limited a pao i nodes from further consideration, Seaary cur caleulatons for each ofthe seo 7 A Flow-Shop Example 108 osha tne meumerer wetter ater pur nl me. This Ihe ose our posse sche we Rare aeadsclminated tele Turning tthe casulaon of (0) for tcemember st. me we sta further sehodules ae eliminate, Comer @= {Jy JI). Here we Eon tht 1 sould scheduled last Thus we move fo 31X10 2134 ‘innate move from LAX to 1324 and from 23XX t0 2814, Sik is re ber pairs of achediesafeehimiated in cael ¥(Q) there cer sete S0 of the ewentyour possible schedules ne have now ‘hevnatd wcmy. The fina calculation 0 FGI, fy) see ay Srnul schedule from te romain four Hence we sce fhe aantage tat dymaric programming ols over camp mera, The tern very pnp tm XXX The ahve no aly aves some insight no the advantages of dynamic progsasming, bur alo point fo rome of ts problem. The method moves Dut ivubeoucs from XX nal econ. ales wing in then aifcrnt pate ofthe Ire atthe time For example. laminates the Ione 43XX fo 4312 by comileaton of the move fom ISX to 1432, Ths simultaneous working puts extern demands on the menor requirement [must remember much inftritton at exh ne every age {noone of hi informationSecomes redundant ithe fl selection ot the optinal pois Branch an bound esos thee in is "nove from XXX ina very uneven fasion exposing sme ranch father tooking at ers aa Senay. continually check 0 ee seme of soted information fas fete ronda ad sa, hs pay orp, ‘Became bh franc and hou and dymic programatng explore he “inition te infeligeily,dctenmining fm rue which Branches ee hw he fly venigate. they are called ipl enumeration metho. Teli ini ogi the checking of every posse seed bat wake ‘ample or explit enumeration They tit sonder every pomsbiity 7. A FLOW.SHOP EXAMPLE We sl icznduc the underyng ideas of branch ardour though & Sipe 4/3 example with rxo ready tes, TH Cou ad Fg a Scualent pertmunce masts for this example The spe: formule thn hat we adopts developed indepemfenty by Mg and Schrage (0863) and by Lari (1968), Note tat hy Theorems §.1 nd $2 we ness ony consider ermtation ue Branch and Bound Mets Ich? schedules Ms the simian nce forthe prem wile a denial branching state to ttf Pig. 1 Contato te ecg of rah tn Round ts nt surprisingly, te des of bounding, Soppome me ae at nade XXX fhe carnuan ce Then shal eae a lower bound {oC the i pol shes tht He Beyond this aod. vi. 2734, 314303314. 3841 a1 and 2481 Comoran of this Bound wie a ‘eter expe twanch of he toe ay Further Bu that 1oKing Shea Fi we mat cie the ours themes then We can EUS Fo notatnal umseneace we flow or convention in Chapter Sand Hobe Pon 3m 6 = Py FOr al jb ts wll so bat gents: our problem to jobs while we develop the fev of the bounds Suppose we ue st 998 epg —- Dag XXX. Thus Kb sp diy Jaap ave Been asgned 10 the BSC K positions in he peossatf sequence, Welt A= age Dre Ju De te uate ‘dosnt of yobs aleady ssgned and U those of Gn —K) jobs as ye Shamed. Fork = 1... K weet ay, be the completion ese of ob py om achive 1 fy, As Simpleian time om machine 2 a jy iS empleton time co michine 2 Bye An gay De calculated ecu htt t tse sa aa, nt a - hunt ay MANE r Bou + Paw (7D _ 1 Pye ty = MBNBe Ta od 2] Ti develop the wer ound we conser thee posites tat aren a serve the mon orale ha ead cr We ivesiale BOW quickly tay complete the nmtiiag fobs Uf e compact te processing 09 cof the hice mcs ur Ft we conn the posh thatthe procesing on machine I in continous: chi of course. always tue for {ny semractive sce. Bal suppose thatthe lst ob in these, sy Mocs ot have fo wait ore acing 28 machine 3 0 be fee ‘fore peng procened on HEE SCE Cy ™ Ce ME BING Can Bust Sat Ci Co “Thus, choosing Jo hae the fas ttl recessing on machines 2 and 3, wee ace tha al sceslen mas have Can ® 9u05~ Ses it #6) 0 [Nest we sonidr the posihy thatthe processing on machine 2s cam Aun. To gg ths we must astme tat thre 0 ee whe a we dls tee 86.7.4) A Flow-Shop Example un sais being processed on machine 1. Further suppose that he {Rijeh the achedule, can compete on machine 3 tho needing fete fede to complet” Then we have Can = Bess SB Ea {Choning Jy 0 ave he east possible processing tne om machine 3. See mata edule must hare oa Finals we suppose jobs process on machines and 2 such chat vere 620 fee for ile time on muhine 3,46. processing on tha machine ma} BE uous, The ne see ha al bcedales mast ave Cm mt Sf oa ‘Now ine ao schedule can do eter than ayo expresions (7.2). (7.3) tou Ga). scedue can do better than the maxim ofthese Thus we Iie he lower bound D(A) ~ manne, fou bin os) nart Sd ‘This lower bo iy be srt may be sry ose th the mike Span a all scheles beginning with the suhsequense A. The teason Simple. In donving the three somponent ofthe Tower bound me have ‘sun that there need not be wing time nore ne a cesta pots inte schol In many ease hs oe no Be 3, ‘Wi these dtintons mind we now tur othe sobton ow pute bar 43/F/Ciu problem. The fable Below gies the processing ties, We "ume tha She ea times ae rere Processing Ties po a rere 2 8 & B 8 gees) u2 Branch and Bound Methods (on? For convenience in calculating mine, + and min (6) fr ‘aroun nts of unsigned foe U. we abe tabulate # a0 | We beaia by cakwatng lower bounds 0° Cau 2 al theres om the Braet the climintin toe leading the processing sence 12344 te te own KNX tr IXXK to 1OXX a finaly e234. Nara at The inal aode 1234 we esate Cy tf and nota lower Bound. THis we bein a a AVX A= O).U = Hadad Myer gee Ane flammild. Bt St ay esi = 9 10,15 621 + 218 10} = 28 API2BE A= (edo Jy 1): there are no unassigned jobs Ba 85629, f= mayo, 3} 721 7S ian[1 5 te} +8 26 2.94.32 12. fy= mai? 18) 2 9= 24 Fon mma26. 26) 42-28 Hence Ci forsshedute 1243, 2% has we find the Cay fa the schedule 1734 28.1234 i our Fst ea sehedute Quester bet epore th elainaton ce compering HE voeee Boundy at cack node wih the wal of Cy, for the cuter til thea te lower hound node pester than eaual to his we rw hat we cannot haprove up the til sede by expening that brant further aad Bence we emia at ads ada aes eon the branch TF the fewer Bound ata nade en thin Cy. Of Me {ehedute, we cane! eliminate the node wn must explore she brane See. 73] ‘A Flow-Shop Example ua “@ nox S a® Fi. 22 Th anh fhe nner i ew it I me ative at final mode and ind that the sche tere ‘Coles than that ofthe tr then thi schedule becomes the new tl scitdule, Eventually we wil have eliminated or explored all the nodes and the tial sbedale that remains st he optimal Soleus return go our problem ame apply thi scene. Fig. 7.2 shows tis one branch of the ehmination tee tat we have explored 3 ft. The rumbers by the nodes are the lower bounds. Now the lower boud st [XX is 2, which se es than Cf the thal Thus e canoe improve ‘upon the tral 1234 by exploring the branch to 1243. Slay the lower rund at nace IXXX ilo 28 and we may eliminate the Branches fo nits 134% and 14XX ang beyond. Thar we have arrived gt the suze fe) te) us Branch and Bound Methods [en [Nest we explore the Branch XXXX to 2XXX stl Hsyond. We make the foo elton AQKKX —A=0).U= pdt WIA) ~ tman(2'© 106 10,6 194 2.10 + 18h im This tomes hound ee than 28 the Cy ofthe tra, Thorlore we mas explore the branches boyoni the ode SEXX. ADI A= 1).U = De BaD l= 3g 2 aa,6) + 8~ 14 y= maait4. 1) 64 18 Bay tai 9 10,184 11+ 2,18 DOH ALIX A= (p10 Uy dh f= 256~ 8p, mani,6) +2 ~ 10 ,=max[l0, 11) + B= 19. BASS mani dy 1,104 17+ 2,294 OF an 240xx 19. = 8,3 DFS, ge mas. 0) + 9 = 15, Hamas, Wye 2 Bays imanid 610,15 4 104 4,19 Fh -"8 “The loner bounds teach of the nes afe notes than the Wale of Cae forthe til schedale so there ran need to expore tess branches wether “This we area the sation shown ig, 7 ‘Our ext tsk fo eaplote the branches beyond XXX. AIKXX A= 09.U~ Urdadd 5268 Bie TA) = mani6 6 = 9,4 21+ 2,164 13} Sie this omer bun greater than 28, we explore no further long tis brane. Finally we tok a the branches beyond 3OO AMAXX A= GNU = Bodo) jaxf3 9912+ 4+ 4 184 17 i Since this lower bow i ater than 2, there eno need to conser the branches beyond XXX risen A Fhow-Shiop Example AS We have now expel the entice iminition tee and shown tat no schedule cum nave aC tse than 2. Sotho cuenta sched, vie [23 optimal The simple elimination eee for the ahov slation i seven in Fa 75 our solution here i happened thatthe fet schedule we evaluate Fad we explored te ranch SOCK to 4321 fst te rl chee would hve changed nthe coure ofthe solution. Soe Probiom 7.61 Ths type of search proceduce we ised above ie cilled 8 depthtrst search, Inter won, our search ofthe tce we seleted branch and Ssitemataty worked down i unl we had either elminted fon the {rounds of lower Bound wha reached inl node, whic ha either scone the ia schedule or was clininated. This seach strategy asthe ‘dvustage thatthe computer need only remember the tay yf 80 bya} fore nodes in he rach corte being seatched Sine theve ate shina (a I) such nodes in an nj problem, this serch procedre Teouies le rage Table 71 below it the nodes at which hese quae ‘ies must te evslsted and remembered ia order to chminae ceria ‘ther ne in the tee ‘Nie that po time was it aoossary f0 remember the value of yf avand IO(A) wt me than 3 4~ J odes. Moveover, because othe ‘Skematic mayo which we searched the re, once the fete eed for these guts at a node as past we Knew we would mover need Them ‘We see then that this dept sate eequites very title storage Branch und Bound! Methods on? gee 73) A Plw-Shop Example a {able 7-1—Memoryroguirements during the depth st search SB Yu ne Hower ads liminated o fully earcd nde ried at des aie IXNX. 12XN. 1208 283 IXKX, 124 XXX and beyond Ixxx DUNK an beyond DANK. JINN 2550x dnd bovond XXX. 2 SIXX and heyond BNxx. AN. xx and tejond SxxX IRN and beyond IRNN fon. however, require 1 great deal of computtion. We were Tacky'—wel, I Set the problem tha the lower founds eliminated roof the branches carly on. I coud Rave happened otherwise, We might have needed 1o explore the branches moch fare Refore fey mere ‘aniate. Se Problems 76. ‘Am altri seareh srtepy 6 frombersareh oF branch-rom: lowestoupd. Ashe later name ges nts esas raneh fom & We withthe cutest lost lover ound. Tee perhaps ease! 1 follow is procedure though an example. Bel’ we se the 43 FIC yy prob lem agin but histime by afonlerseach. Phe cleans teading to he liner hounds cher have ben gen area ar ar et to you ar exerci. Fist of al we simultaneously ranch to each ofthe for ees TOK 2XNX. XXX. 4XXX and alate the lower bounds there. Thos we ‘bin the rest shown a Fig 74 The made 3XXX ha the lest lower ‘ound so we branch fom ths wo nodes 21XX, 23XX. ond 24K to cba the esl show ia Fig 7-7. Here we fir! two modes IXXX an D1NX, ton? sare the same lower bound. Peking 21% abiteay we tan co the fndes 2184 and 2143 46 shown it Fig 7 The schedule 2138 bye oe = 28, Thine 16 the lowest af the lower bounds oa athe eco Thay 2134 5 an optimal chee (NB schedules 7134 and 1234 ate both optial with Cy ~ 28. Had we branched at [XXX latend of 2UXX the frotiescrch would have Found the sae optimal schedule asthe depthst search) Hs the node 2134 not usned out ob clearly an optimal sched, we woul ave proceed follow. Ether schedule 212497 2143 would Baye ‘nxn selected us the tal according to which has the smaller competion time Then those nodes with hiwer bound renter than tis est et Con ‘would have been ekminted trom further scurching, Ths eine done, twnching wuld hare Continued from remaining node with the least Ter hound Continuing inthe vious vay, an optimal shedule would srcntualy ie found sec.74] Pains about the Bach and Bound Method ny tn this example, the frontier search des ot find an opin shee sane than the depth fist seatch, However, wee Problem 7.1 There ou Sreashed to soe the abuve problem with a depttie serch tht seeps STowethe Wee fea eight to et rather than fom fe ht You wi find {Fidos 0 thar you need move clelsto than in the above Foner Starch la geneal frontier seateh wil recite lets salclaton tha pest Search "The fowtier search choos which branch to expan tent a more itligent fashion snd so wally ind 0 opal Slaton finer Unortunatly ths beter performances bought at oot of greater Sorapefequvements Conti the menoryseauemets needed inthe neve Here we are required 10 keep ta0k of 3 By iy 898 the Hower oun as amy a8 7 nodes, The dept Search only requ hese {uunitice st mow three nodes at ime Tube 7.2 shold ale ake ear the reawn for allng the branch-from-lowcst bound senrch frome search. At sue stage i cmst remember the hiner bounds of Ihe now ‘iminated odes onthe frontier of the nes explore! Before eksing this scion. noice atthe abone example does aot seaform the caftions (5.2) under which Jason's hay may he applic we apply the arth nceles, we ebnan the chet 251d whieh by the lamer hound a¢ 23NX tse have Cn = 29 and anaot Be tinal tm lea thrcfone that comdton® (82) ae mo ‘redundant, But are ascisary fhe sucesso that algerie Table 7.2--Memey requirements daring the frontier “Promter of we at which Ban Brmnehig yan the loner howe ae egal 1s DOO DAN SHH ANAK mms INXX INK 25XN MXN. NAN XX int INRK HSV 2183, 3KN SKK XXX AXXX, 14 SOME GENERAL POINTS AROUT THE BRANCH AND BOUND APPROACH Banc and bowl is Bech used to aulve 0 many cles of schdling problems ht deseres tbe Heated save genera. Thus ne sow de We togin fy studing the Logical store uneling st seaceh of the imimation tee, Indeed ir fet ue sonsier the construc of thee aio tee cl 120 Banc and Bown Methods “The fist requicemeat that we sbouk be ale to repeatedly partition the set al pl schedules into emer an ses subsets unt a yan we obain pactinon of 2 mtosobete whch ese conta one and ony ‘me wheat. Thus. if 2 5 st parttioned into... 2, and then ch is 0, f,,-- and won, we obtain the we state show Fy 79. "As above we cll each substi hse @ node Inthe example of he las seston st may not have Been obsous tht nodes corresponded 6 subsets of ecedules. bu they dd. For iastapce the subst labelled, TAX und was the subset of chedues wit od 1 fit inthe procesing ‘Sequence Simsy fs nas labled 23XX and was the suet of achedales sith jo 2 fr an jo 3 second ia the prcesaing sequence AR s0 om ‘Apart fn thc branching strcture of subsers th method ao ries tate have Bwana Thus isthe subvet cmrexponding. toapottuus roe, weston able a calculate a lower Bound 00 Ihe performance mcinate fr all the schedules in Yeo the value of th permance courte fo 4 sched v then we regu. thereore, [nr BQN =e(y) forall yin. ‘We need tw further components to define a geeca ranch and bound procede. Fry theve i the ached, ¢* Th my 90420 ial ht st some post the proce Recomes tf the hed sth was the best salve ofthe pettormarce mewure found 80 fat Set Truly we need a search ster. Wich, aot surprisingly. fells us Bow 9 ‘Consider how a branch and bound procedure progresses wards a solution At any stge the nodes 2 prone int re cases 74) Paty about the Branch sa Bound Meth m1 the clawsof nes which have ether een eliminated or fly explored. 1 fly explored ose tera termined ode for which the perfor: mance mesvae as been evluated tact cra de ia he a hoy ‘tthe tice sch tht lower bounds have cen ested a al te ses Immediately hevond Zine aso partly explored nos. The lower boul bas been eva sed for any node i his elas, But the aode shell hay heen eer ‘Siminated nor fully emored. [Une cass of unenpored nodes These nodes neither have been clin ted pli th elimination of» node tat proceeds her the tee nor have ben examined 0 far bythe setch posed For instance, in the frontier seatch solution of the presi section athe sage cslnyed in Fig. 77 the sets 2 Ze and Zy ae 2) = (RRNX, 2004); 25> (IXMX, INN, ZINN, 28NX, INNX, ANAK othe remaiang 38 odes ale in 2; "A branch i aid tobe fated when there fw freer ees «0 seatch along H. "The happens in ane of tne ways Ether a node and ence” te subsequent noes ate eliminated or the tore mode ‘ated and the perfomance messue evaluated there At the hegining of Bach and Bound Slution al the nodes i sess Aste soluben progresses nodes ae moved iter del toes Zicr ig the imermediate uss 2. The plocedre ends with al the odes inclas2,,all branches fahomed and the knowledge that he carte ti) ‘eh optima ‘We now examine stage of whranch ad bound solution tose exctly how mades are moved ftom Zino Z The seach procedure selects a pit faces Yana Y, such that ¥, les dey beyond Yim the tee and Imersver. Yies in Zand Yn 2 In oter word the tree Branches fom, Inde Yonodes YY... ¥ isone of those absequent noes. which Taster io be explored. Tete tas dil when we cansiger te fist ste ntl al he noses lien ZZ, empty. Thus he seth proce. te wi og be able wo sect sable node To avoid tis problem he bis stages convention one. which he inital node 9 Frese ‘he set ofall pombe scheles i placed in 2, When aa tave been cesta thee ate se poses’ iter Y ‘rss more thn ane schedule oY, species a sehedle uniquely, 2 rf. We cond ea of these posaliis in ur, 2 comains mone thin one schedule, We calelte (¥). IF ‘cry =e") bet vai ofthe performance masa ye then We eliminate made Fad alt the de that Besa Tis Mey ate m2 Branch stl Bound Methonis ton moved deel rs 70% HIB) <4"). we same imine Y, in 0 mes how to Y,= bhi comin esate one whee, Hote we ate fe). etsy me(er). we shminate Find this mune than Zy 86 Zh E13) = ets") or te ral sole Bas mt eS hen Bee ‘hemo ial chute In thiscaee me mun examine lhe ales to se if thy te clined y he vale ofthe peormance ease for thi es a schodute.Irsurh 9 nade feiminated, He me together wih al the andes beyond tn Z te eer of the shows cuss the examination of ¥, may some the plata ofthe oe ¥ Wr we mone ade fm 007 ile hough shows ht forum sonable ere ste the nc bound preedie wil evenaty trate wa he Roden ya {he ral sos rensane l e ope Inte ses we met arches he pti nd he froner The apf utegy sce ae F the mae st oy pled inZ. (orth fason eal caked the mowest-actveaode serch) Ta totgy wl aly nlc as the ¥, ote pcs ae ho progr Singemtey node downs branch However te revi sage hed {re Branch he sare retreats yp the ranch ani cree node tha hw ot Ben fly explored her searehen ore of he unohomed tranches eyo hs nove he fronts seach egy more spe fated, When completely feet share, i selets 9 4 Hoge m2 wih fas he let ower bound. Tis coi fies ¥ un thas bot completely ‘splined Tho the ie branches tom Y he susequent nodes Ye Pa Yothe ext sage othe aur oe Hed Vy Yar Ysa comin "We have sce that dynamic programming hast major fas. Fit requires tat very many quan Be remembered Caring the slo ‘Second though mach aster han complete enumeration hy sone. Tes tke probly long fo soe medio age pens. 1 brane tnd wad uty othe sane Umaeatger? We ome Mage ‘quirement ew Sie miu sve of 7, determines the mags rogues of & tach nbd peor Frnt itr ata ny 6 {he precede mun seme ocer hounds ober terse a {Wat st and ly those ns con ts claw. Ne qua wl ave been caused for any ne wheres the wet tous ad Cher seis atthe nos in aloo elated al ane Sid thes and wil ave Ren fog Ac we cathe a Sct inn depthat arch wi never berger than the mer of ees nth ps ranch the lirtion te. Ths emacs mack OE Seo 74] Paints about the Branch and Bound Method 123, than the sore neguiemonts tan evant dpa graoning e-Fowa frotie natch tho we of Zh compete anpredcate, #4 Icors the we wf Zs osm extras np daring the sono sine pve Te practice. however th de mo scorn Bape ad Frog sarc mat hea eth ny vl sk ht hey el exbast fae avaiable storage copy “The sumber of pratons required and hone the time reguied 12 suse problem by french and bound iv amprediciabi, whatever sate iateq bse, It night Rappen tat the pons has tn enone fll Sirwally every nde, which cae st would take ts long complete ‘umeration. Indeed. it might take longer Because branch tad bond insihes more computation “per nage thas eompicte enumeration Nonetheless n general brinch nd bourses peor 3 great eal betes Tham comple enumeration. But W shoul mt be serie fo ti ae ‘Se sue any problem io practice. Thott lke Jame propeans ying aways finds an optimal solution, ut ay take probes Don. {hs For instance 10) LG/8 problem posed by Muh and Thm ‘se (1968) sll has 6 e solved optimally (Graal. 1970) ‘Obviously she search stategy sed isan important deterinan’ ofthe tine requited to solve 2 problem. AS we have alleady noted. a omer Scich generally fins a opal sede aster tha 3 epi search ql iiportns the quality af the lower bounds fhe hone bound at ‘ume ts god te, not mich less han the et value of the performance heasure there then the procedure general’ Finds an opin schedule ice examining fewcr nodes than f would sith pooter bounds. Good founds eliminate nos high up the te. thus reducing he seatch sub Sfanally. Conse te exarpe ofthe previous section, Ateach code 210 isalower bound abe a mech poorer one than {7.5} Wisi 2s wed ser branch must fe explored Cleary the etter Bound (7.3) to be Prsferre, Despite these comments must not be assume that the Beer the lower founds the faster wil branch and. Bound find th optimal ‘schedule. It depends ypon now long the eaculation takes am each noe. ny be fster ever caleulate poor ler Bounds quickly ad examine ‘many nade inthe earch dha to ences god loc bounds ally and ‘ouine far fever andos, However empires invesgatons huve shown ‘hat generally the extra fort ruled to ciculate yood lower bounds tise wel spent the rat of eiminting Hodes nceass ote tha enough ‘wsompensate. For instance, Baker (1975) has compares the perormance ‘or branch and Bound agorithm for njm/P Fy problems based spon tounde similar 18 2.5} with one based apon a Tee compa bound of NeMahon and Batten (1967). (See Problem 7.66) He found tht for SSnpie problems the enra calcula for she second bound wat not ‘worl, Dut asthe poems became moe ifct he second proce- wa ‘Senadh ond Remedi Icha arbi a vo lh stemmed Rimny Ranta (173 re tae gre inub item par of he chen wes, Wie aaa ght tor actninacd hoy ssn nsvare nhac fae kero ih te sch oie bs loner tpapent soe to fs Sim ng ur tt pee igh fe spend ig ‘Stig vp tous the per ean gt bap fom as ower et : extol woe ay aay ae endo og ak Sma ue Psy, ay pone cee Steere Thelin thin Tne toderin kn ower tore at Sips tobe Sindh a ages nfo acum on Shuster nin shee emp ae ounces mech So Th sanpuon of coe Sogou! ont nite tearm: ya fa war api ee? 7 rer Songs Ws dine eg ea Sat fbr yw ate amit erates wt sn Sreype om crest tone ec: pod ot opi, SHSMBS, Princes Snowe tm on § ts oped wean Pes yt rah ands (82) a0, en Stott fl scl obs hs eth i me sen al chs fos anc anda ts, Bo Ber Gina arate (978) tpn oie ag Shes wit ao tl hee cn won se Siiaup te mys ones Ss, We hae sty mgd tn ym owantng enh Tatoos ms ater esque Sop we vest contr ‘htt wnty onions omen wey Stes a SRtunls ome mak doo tern essed ate ee ‘Tor choy oe oye he st mde ene ost Fr example, comer an /97F ple and ree othe nota Sie ea Bee TL eons eon yee ARR Dural Ke eae Whe ans oO OK pol tae Sri ete pe sare sone ed be Sy incr he ig ins science tnnemeinnasenten scm cm Boo Baer Ya > Yi ithe completion time of he Kh job on each machine i eater Under theft ubmequence tha itis uner the tecdad. Then, owever we com plete the processing of he unassigned ube the ist made me sal do Kast aswell with the same completion at the secon Ths we ay i> see. 75] ‘References and Further Reading Bs rats the firs nade from the search, Chssking for such sorinanee ond Inns during the search ofa tee may tke consilerable computation 2 Sova. NonetReles. mee they may climate mat nodes before Tower ‘undingargamene can dos their use may cura the seach uit that overall redaction sn compaltinal telremente 16 obtained leveed this hasten found in pretice, Baker (0975). Rinnooy Kane (C575) and agowege a (1978) a report tha atta iclason of dom hance sors branch and tnd can ead to improvements pr 2 final mete for ving computation ew accep ah-optinal 0 vn Sappone we ace tht sy solaton within 1D of te opal be suenfaciony ef jq is am optima schedule and tv seele soch thate(s) = 1 1e(s,h ahem me would accept asa soln. Now sippose tata some stage Ska ranch an hound serch we have onda Schedule" Then we may eliminate af! oes wth ower fous pester ‘tan eU")/1.10. For ioe of those noes were Yo lea tan opin tthe, then it wo Rave Yo be win 10% of our eurent tal Ths Siren looking fora suboptimal solation we may eliminate nodes later than t'we were only prepared to accept ap optimal schedule. Kohler and Seite (1998) ror that the savings computation that res ny Be seat 17.5 REFERENCES AND FURTHER READING Conway, Mawel, an Miler (1967), and Baker (1978) both comin tnoductions Ye the we of branch and hound it scheduling at & lave, Smparable sth ou? presentation. Lensra (1977) und Rieooy Kan (157) dhows the sbjet in mach gree technical deta As remarked arr branch ad bound methods have Bee tbe most succusfl oF the nonseunsic approaches for solving sebeuulng b+ Tem although there are ocesione when a dynamic programming soiion ‘sine prefered, eg sce Baker anc Schrage (19TH), Given th succes Insy seem strange that we Rave orl studied 2 branch and bound solution of one problem. namely te n/2/F/Fyq one. The reason is simple. We ‘ed thatthe ality ofthe lower boas in a importa determina of frkuemance of procedure. Hence. mac sl as gone ino the dei of ‘hse: Ta consequence. the resting hounds te based upon mathematica {acory far beyond the level of that noamed i ths book. The following Feteences seus three important appliatons of Draehsn0 Bound 1 ‘tung but they do equte rach mathematical koomledge td ls om the part of the fender. ‘Rinaooy Kan eta (1978) comidr the 1113p robles Lagewegr eal (1977) consider th mln/GIC problem nd Lagemexer al (1978) the most uptodate eference on he peta 126 ‘Branch and Bound Methods. [chy ion flow-shop problem Pye thet theory superceding that deus Sed in Section 73 (ge al Bary 1079; and Pots, 1978) Burkard {1980) dncuses a yneral approach Ws devclpingKower bout for job shop problems Tn Seton 7.4 we syne the general struct of branch and box, and ako Row we might improve the speed sith whi a solution Is found [Kobler and Step gine a mach more dened and advanced restnent of ‘hese dens, sapporting thei diesen wth bo mpi an theoretic ‘evils The themctio performance of branch bound is abo Sneed, by Lanstrs ant Ruano Kan (1974) 16 PROBLEMS 1. Soke the exami of Seton 7.3 bya dep isfit earh that seep from right lf i tart with the branch fom XXXX to 8320 2 Sole te fluming4/34°C,. poem wth ero ready es (by sept serch, (8) by a Fomor ear Processing Times Job Woe ww ease 2 ow § & ¢ ob Ww 5 i ew 3, plain why she method of Section 7.3 canna be enended to ve ‘/4)F/Con probleme ten the method to sole the 44)P/C problem Sth zeny mcs and prong mes soe See Tat Problems nr 5. tered are he processing ins hth machine mt de Stes, "iL Extend the method ot Seen 7 tale for mon sea fe Soke the m/3/P7C ble th at Coal yo extn your method t9 whe an H/VF/C a, pele with non rem rowdy tes? In Seton 7.3 we wed the kines hound 4A) = masta tet + Sov emaths sen +S mites not Se Show that this bon aye ipod D(A) ~ masta aa 4 = mania z,~pliptedt = minted > masta Ba le) intl ras ‘Use thisformofthe joertound lose the 43/6, obo with dat: 18 [cnt Hl = a mo 8 4 ie 1 3 6, Show thatthe folowing eso «hower Bound the Po yy Jy ‘yar BX for the nH4)o, prem, VAY = nas tana +, 665+ Sintec} 7, Conse the (Cy problem with sequence depsndent poseving limes ak dsusned in Prater 69.7 Develop a branch snd Pour ution is fotos Nice Mat he emination re wl hae the same te that duns in Section 7.2. At woe, Duar XX NOW a tower Pound DAY ry Raves tie Pa Gee) where og =n Hence soe the 4/1/(Cy problem in Prohlem 6:5 {Sind = suboptimal slut within USM OF she opis! Cor be (9/h/C 2, problem wit dat sw 761 Problems 129) ‘Yow may ing ise t your ist ia sation 1 that whieh Joon goitim would determine Wt was apptcable Solve the 9/17 probe daca in Seco 6:2 by bench ard boon Chapter # Integer Programming Formulations 1 INTRODUCTION There is 4 sizeable Hody of erature which suggests solving sched problems by recasting them at mathemie! programmes parte Imeger programmes. (A mathematical programme i sinply general of eonstained optimisation problem. See below) These teas probl may be solved by standard algortims, which have been developed Solving general mathemati! programmes. Hence, tanlting back, ‘btan opti Schedles. Sat simply, ths sounds avery promi approach in practice it ao. Te standard mathematical program lgortums ae practealy spicale oly to small prublem the re ‘eheduling problems can be very are ined, Thus the cain se Just thatthe inherent dilfieaties are rephrased, bt no non ore ‘ble form. Empirically this confirms that chedling permeate ‘rl very difcalt~and donot jst appear tobe 0. Tn the next section we translate an/m/P/C scheduling problem i 1 mized integer programme. We shall ot cscst methods savings problems, but Instead refer tothe erature. In the seme that Ih Indicate, such reformulations are misleading since the integer progta tes are no easier to solve than the onginal probes. Noneteles inoductory teeatment of scheduling mould be complete without ‘mention ofthe following. Moreover, the modeling tchnigues we are considerable importance eiewher in operation research, “The body of theory and algorithms called mathematica! program eas with problems of the flowing Kind Minimise fly x5 4) with respect to), soos othe constraints sec. 8.1] Introduction 131 Btn td by Edtntn cad aby ent A=, In other words, mithematica programming isa family of techniques foc optimising fancion subject to eonrsint spon the independent r= thles In scheduling we wish 0 optim a performance measure subject technolgieal conranton the alowable processing order. Thus ts not {erpsing that given a lie ingenuity the ater ca be tensed int the former type of problem; both concer opsimiaton under constraints, 1 imeger programming oo be sicily corest mised integer pro ramming neo the inde nen variables at conslned oe negra ‘ten they are only allowed the valves Oo Lan are used to inde the sees or presence of some propery. Furthermore the functions 8 ave Imear Tha the stanard problem take he fom minimise ry Haar, Feyty te) subject gua + Batt Reet teh Bot, kan Rant Rye Shh Bots Hats Hgts + BUH nd certsin of he xx ae limite to intgel values Not as th time of the inequity constraints may be replaced by srct equality Methods of solving such problems ate reviewed in Gatinkel and ‘Nembawser (1972) and mot introductory texts on operational resarch ‘¢g:Daellenbach and Georg (1979). By and larg, those methods can be “lasified aseither impli enumeration, in partial branch and bound, oF utting pane. Both requte mich computation. Moreover, both are based pom property of nea rogaine gn ad py ear 0 parcular properses ofthe problem being saved. AS a esa they tend 0 take longer fo find olution than implicit enumeration algorthns \ksgned specifically for parla las of problems. For example, im Integer Programming Formulations [ch 8 sider 0 methods of soving a scheduling problem. Fin, me may Tate into a nese programme and solve hat by branch and bound the bounds tase upon general integer programming theory OF, secondly wwe may tacle the problem directly by branch sd ound with bow based upon our knowledge ifthe physksl proper of schedules, lower bounds found in the fist exe ae usally pore than those found te scond ease andthe branch and ound serch coresprading Tonge So, we have sid, sis better to approach scheduling pubes di rather than increcty via inteper programming. Assumprlon for Chapter There reno specie estrctions up the aplication of integer prog ming method. 52 WAGNER'S INTEGER PROGRAMMING FORM OF n/m/P/Co PROBLEMS, Wager (1959) hes introduced the following integer programming lon of he penton lwp pri im) P/E ar Wess allready tines ee ze, AAS sual he subscript will refer to the jb 3 the subscript 10 machine Mz and the subscrpc 10 the Ath poston inthe prose sequence. Ths the jobs wil be proceed thigh each machine ‘order hyp Jute <--> Jy =o Jn). Moteoner since the petmut ow sop isa spesialistion ofthe lwshop, each jo has technol constraints (My My =. M.) To model this problem We introduce! arabes constrained 1 tl the values 0 off 1, if iseheduled ia the Ath postion ofthe processing sequence Xo Jo, omcrwise ‘he constrains that must be obeyed by thee variables a Emad for bette ie exactly one job i scheduled inthe kth position; and Sxe21 for i t2en {cach job x scheduled in exactly one positon, These constrains ach X, to tak the values 0 and 1 only, provided me alo demand that Xj are nonnegatine integers Next ne introduce two sof nonnegative rel variables; [and Wy Sec. 8.2} Integer Programming Form of n/m/P/Cu.,Problems = 133 The fst st Js opresea the idle tines onthe machines; henson se, Wrthe wating times forthe bs betnen machines The tne faba here wil be shy diferent to that intoduced in Secon 1S. Specialy 1's the sale time of machine M, between the completion of jn the Prosesing seauence andthe Star Of yy THM les ened for) 1 2. im and = 120021. Morne, sine hee need be no tine onthe fst machine, = Ofor k= 1,20 n= 1 The malting tiene Wi the time that Jy, must spend Betneen completion on Mand staring 'o be processed of Mj. Ths My fs defined for = 1,2, Im Vand ~1,2,-+.n-Moreover, since ther thing delay the fit jb procened, W, =O for} = Ir de..m ~ 1 These quantities are Indicated in Fig. BA. Noe hat, and W; cannot both be now 20 fora semi-active wheal; they ar thom ch nthe diagram purely for “The nea sof constrains hat we inkoduce simply ay that hemes 2, erween completion ofthe jo Jj0% M andthe tar of he Jy. 00 ‘ies ost be mel defined From Pig. 81 fear that Te + Daas Wines Wa + Payor Base es) ‘Weems expres pa. ahd a im terms ofthe Xq, Because we donot now explcily how to find the subscripts (+1) and 14). Now, since Xin fezer except fori = i(k + I) mhen itis follows tit Exar a) REI The esis Seen teas te ieee

Vous aimerez peut-être aussi