httpsoaities
2 Flowtime heuristics Algorithm RPI Time Type PARETO Algorithm RPI Time Type PARETO
1 PR1(15) 033 2093 Composite 0 20 RZ-LW 129 1038 Simple 1
2 PR2(15) 036 1756 Composite 0 21 IH7-FL 130 8606 Composite 18
3 PR1(10) 039 2012 Composite 1 22 IH7 143 8745 Composite 19
4 PR2(10) 041 1702 Composite 1 23 C1-FL 172 6398 Composite 17
5 PR4(15) 041 1664 Composite 0 24 LR-NEH(15) 172 494 Simple 0
6 PR3(15) 045 1743 Composite 2 25 LR-NEH(10) 175 330 Simple 0
7 PR4(10) 045 1584 Composite 0 26 LR-NEH(5) 184 165 Simple 0
8 PR3(10) 046 1731 Composite 3 27 FL 199 6124 Simple 20
9 PR1(5) 050 1753 Composite 5 28 LR(n) 209 13412 Simple 26
10 PR3(5) 051 1458 Composite 0 29 LR(nm) 229 681 Simple 3
11 PR2(5) 051 1404 Composite 0 30 RZ 265 094 Simple 0
12 PR4(5) 053 1332 Composite 0 31 WY 283 4190 Simple 22
13 IC3 062 7725 Composite 12 32 LR(1) 313 029 Simple 0
14 IC2 066 1995 Composite 10 33 NEH 403 037 Simple 1
15 IC1 081 1441 Composite 2 34 Raj 502 008 Simple 0
16 C2-FL 095 13822 Composite 15 35 LIT 826 142 Simple 4
17 LR-FPE 114 1035 Composite 0 36 SPD2 1656 143 Simple 5
18 FL-LS 122 12024 Simple 16 37 SPD1 1737 118 Simple 4
19 LR-BPE 123 1134 Composite 1
httpsoaities
2 Flowtime heuristics
PR1(15)PR2(15)PR1(10)PR2(10)PR4(15)PR3(15)PR4(10)PR3(10)PR1(5)PR3(5)PR2(5)PR4(5)
ICH3
ICH2ICH1
C2_FL
LR_FPE
FL_LS
LR_BPERZ_LW
IH7_FLIH7
C1_FL
LR_NEH(15)LR_NEH(10)LR_NEH(5)
FL
LR(n)
LR(nm)RZ
WY
LR(1)NEHRajLIT SPD2 SPD1
0
20
40
60
80
100
120
140
0 2 4 6 8 10 12 14 16 18
CPU
Tim
e
RPI
httpsoaities
2 Flowtime heuristicsPR1(15)
PR2(15)
PR1(10)
PR2(10)PR4(15)
PR3(15)
PR4(10)
PR3(10)
PR1(5)
PR3(5)
PR2(5)PR4(5)
IC2
IC1
LR-FPELR-BPE
RZ-LW
LR-NEH(15)
LR-NEH(10)
LR-NEH(5)
LR(nm)
RZLR(1) NEH Raj
0
5
10
15
20
0 1 2 3 4 5
CPU
Tim
e
RPI
httpsoaities
2 Flowtime heuristicsPR
1(15
)
PR2(
15)
PR1(
10)
PR2(
10)
PR4(
15)
PR3(
15)
PR4(
10)
PR3(
10)
PR1(
5)
PR3(
5)
PR2(
5)
PR4(
5)
IC3
IC2
IC1
C2-
FL
LR-F
PE
FL-L
S
LR-B
PE
RZ-
LW
IH7-
FL IH7
C1-
FL
LR-N
EH(1
5)
LR-N
EH(1
0)
LR-N
EH(5
)
FL
LR(n
)
LR(n
m)
RZ
WY
025
05
075
1
125
15
175
2
225
25
275
3
RPI
httpsoaities
bull Our presented methods range from fast andhigh performance simple heuristics LR-NEH(x)to slower but with state-of-the-art performance
bull PR1(5) results in 05 deviation from bestknown solutions in less than 18 seconds onaverage (less than 3 seconds for instances upto 100 jobs and 20 machines)
2 Flowtime heuristics
httpsoaities
bull Can we improve the results further
bull Most proposed methods are fairly complex
bull To propose simple and easy to implement newstate-of-the-art methods for the problem
bull To carry out a comprehensive computationalstudy for the problem
3 Flowtime local search
httpsoaities
bull We are interested in simple approaches
bull Iterated local Search (ILS) of Lourenccedilo et al(2003) and Iterated Greedy (IG) of Ruiz andStuumltzle (2007) have shown competitiveresults in similar settings
bull We also present population extensions ofboth approaches
3 Flowtime local search
httpsoaities
3 Flowtime local search
procedure ILS ionitialSolutGenerateInlarr0π
)( 0ππ hLocalSearclarr Local search repeat
)( ππ onPerturbatilarr Perturbation of the local optimum )( ππ hLocalSearclarr Local search
)( πππ erionAcceptcritlarr Decide if new solution replaces the incumbent until termination criterion met
end
procedure ILS
Local search
repeat
Perturbation of the local optimum
Local search
Decide if new solution replaces the incumbent
until termination criterion met
end
oleObject2bin
image3wmf
)
(
p
p
on
Perturbati
not
oleObject3bin
image4wmf
)
(
p
p
h
LocalSearc
not
oleObject4bin
image5wmf
)
(
p
p
p
erion
Acceptcrit
not
oleObject5bin
image1wmf
ion
itialSolut
GenerateIn
not
0
p
oleObject1bin
image2wmf
)
(
0
p
p
h
LocalSearc
not
httpsoaities
bull INITIAL SOLUTION We employ the veryefficient heuristic LR(nm) of Li et al (2009)and Zhang et al (2009) This is an extensionof the LR(x) heuristic of Liu and Reeves(2001)
bull LOCAL SEARCH We employ the RZprocedure of Rajendran and Ziegler (1997)With accelerations and to optimality iRZ
3 Flowtime local search
httpsoaities
bull PERTURBATION A given number γ of randominsertion moves
bull ACCEPTANCE CRITERION Simulatedannealing-like with constant temperaturefactor λ as done by Stuumltzle (1998) or Ruizand Stuumltzle (2007 2008) and others
3 Flowtime local search
httpsoaities
3 Flowtime local searchprocedure the presented ILS algorithm
Set the parameters γ and λ
)(0 mnLRlarrπ Generate an initial solution
)( 0ππ iRZlarr Local search
ππ larr Best solution found so far repeat
)( ππ onPerturbatilarr Perturbation of the local optimum )( ππ iRZlarr Local search
if sumsum lt )()( ππ jj CC then Acceptance criterion
ππ larr Accept if better than incumbent if sumsum lt )()( ππ jj CC then check if new best solution
ππ larr endif
elseif ))()(exp()( eTemperaturCCrand jj sumsum minusle ππ then
ππ larr Simulated annealing acceptance criterion endif
until termination criterion is met end
procedure the presented ILS algorithm
Set the parameters and Generate an initial solution Local search Best solution found so far repeat
Perturbation of the local optimum Local search if then Acceptance criterion Accept if better than incumbent if then check if new best solution endif
elseif then Simulated annealing acceptance criterion endif
until termination criterion is met
end
oleObject2bin
image3wmf
)
(
0
m
n
LR
not
p
oleObject3bin
image4wmf
)
(
0
p
p
iRZ
not
oleObject4bin
image5wmf
p
p
not
oleObject5bin
image6wmf
)
(
p
p
on
Perturbati
not
oleObject6bin
image7wmf
)
(
p
p
iRZ
not
oleObject7bin
image8wmf
aring
aring
lt
)
(
)
(
p
p
j
j
C
C
oleObject8bin
image9wmf
p
p
not
oleObject9bin
image10wmf
aring
aring
lt
)
(
)
(
p
p
j
j
C
C
oleObject10bin
image11wmf
p
p
not
oleObject11bin
image12wmf
)
)
(
)
(
exp(
)
(
e
Temperatur
C
C
rand
j
j
aring
aring
-
pound
p
p
oleObject12bin
image13wmf
p
p
not
oleObject13bin
image1wmf
g
oleObject1bin
image2wmf
l
httpsoaities
3 Flowtime local search
procedure IGA tionitialSouluGenerateInlarr0π Generate an initial solution
)( 0ππ hLocalSearclarr Local search repeat
)(_ ππ onConstructinDestructiolarr Destruction and construction )( ππ hLocalSearclarr Local search
)( πππ erionAcceptcritlarr Decide if new solution replaces the incumbent until termination criterion is met
end
procedure IGA
Generate an initial solution
Local search
repeat
Destruction and construction
Local search
Decide if new solution replaces the incumbent
until termination criterion is met
end
oleObject2bin
image3wmf
)
(
_
p
p
on
Constructi
n
Destructio
not
oleObject3bin
image4wmf
)
(
p
p
h
LocalSearc
not
oleObject4bin
image5wmf
)
(
p
p
p
erion
Acceptcrit
not
oleObject5bin
image1wmf
tion
itialSoulu
GenerateIn
not
0
p
oleObject1bin
image2wmf
)
(
0
p
p
h
LocalSearc
not
httpsoaities
bull INITIAL SOLUTION LOCAL SEARCH andACCEPTANCE CRITERION do not change
bull DESTRUCTION_CONSTRUCTION Identical toRuiz and Stuumltzle (2007) d jobs are randomlyselected and extracted Then they areinserted one by one in all positions of thesequence (as in the NEH heuristic)
3 Flowtime local search
httpsoaities
bull POPULATION VARIANTS pILS and pIGA
bull Initial population of size x using LR(x)
bull At each generation one solution from thepopulation is selected
bull 50 of the times using binary tournamentwith flowtime value
bull 50 of the times using inverted binarytournament with the age of the solution
3 Flowtime local search
httpsoaities
bull GENERATIONAL SCHEME
bull A new solution replaces the worst solution ofthe population iff
1 Flowtime value better than that of the worst
2 There is no other identical solution in thepopulation (permutation)
3 The average diversity value of the populationdoes not decrease (Pan and Ruiz 2011)
3 Flowtime local search
httpsoaities
bull DIVERSITY CALCULATION
3 Flowtime local search
[ ]
=times
nnnn
n
n
nnji
21
22212
12111
φφφ
φφφφφφ
φ
[ ]
minus
minusminus
=times
21
212
121
nn
n
n
nnjj
λλ
λλλλ
λ
sequence theof position at appears job timesofnumber
ijji =φ
jjjj
prime
=
jobafter y inmediatel appears job timesofnumber λ
httpsoaities
bull α and β are the number of elements that arelarger than zero in both matricesrespectively
bull Then the diversity is calculated as
3 Flowtime local search
2)11min()1(
)1()1min(
minusminustimesminus
minusminus+
minustimesminus
=xnn
nxnn
ndiv βα
httpsoaities
bull We test 16 different algorithms1 Discrete Differential Evolution DDERLS of of Pan et al (2008)
2 Iterated Greedy IGRLS of Pan et al (2008)
3 Estimation of Distribution EDAJ of Jarboui et al (2009)
4 Variable neighborhood search VNSJ of Jarboui et al (2009)
5 Iterated local search ILSD of Dong et al (2009)
6 Hybrid genetic algorithm HGAT1 of Tseng and Lin (2009)
7 Hybrid genetic algorithm HGAZ of Zhang et al (2009)
8 Hybrid genetic algorithm HGAT2 of Tseng and Lin (2010)
9 Genetic Local Search AGA of Xu et al (2011)
10 Hybrid Discrete Differential Evolution hDDE of Tasgetiren et al (2011)
11 Discrete Ant Bee Colony DABC of Tasgetiren et al (2011)
12 Iterated Greedy SLS of Dubois-Lacoste et al (2011)
13-16 Proposed ILS IGA pILS and pIGA
3 Flowtime local search
httpsoaities
bull 120 instances of Taillard (1993)
bull n=20 50 100timesm=51020
bull n=200timesm=1020 and n=500timesm=20
bull We add 30 more instances with the missingn=200timesm=5 and n=500timesm=510
bull Average relative percentage deviation from bestsolution known as a response measure (AVRPD)
bull All algorithms coded in C++ and run 5 times oneach instance during milliseconds elapsedCPU time on Intel XEON E5420 processors runningat 25 GHz and with 16 GB of RAM memory
3 Flowtime local search
mnt ρ=
httpsoaities
3 Flowtime local searchρ=30
IGRLS DDERLS EDAJ VNSJ ILSD HGAT1 HGAZ HGAT2
AVRPD 039 039 772 488 049 223 074 529
AGA hDDE DABC SLS IGA pIGA ILS pILSAVRPD 087 064 083 041 024 028 025 031
ρ=60IGRLS DDERLS EDAJ VNSJ ILSD HGAT1 HGAZ HGAT2
AVRPD 036 036 702 439 049 213 063 450
AGA hDDE DABC SLS IGA pIGA ILS pILSAVRPD 078 060 076 041 024 027 025 030
ρ=90IGRLS DDERLS EDAJ VNSJ ILSD HGAT1 HGAZ HGAT2
AVRPD 035 034 664 417 050 209 059 409
AGA hDDE DABC SLS IGA pIGA ILS pILSAVRPD 072 058 074 040 024 027 025 029
42 Better
httpsoaities
3 Flowtime local searchInteractions and 950 Percent Tukey HSD Intervals
015
035
055
075
095
AVR
PD
IGAILS
pIGApILS
DDERLS
IGRLS
SLSILSD
hDDEHGAZ
DABCAGA
ρ
30
60
90
httpsoaities
bull We have studied the permutation flowshopscheduling problem with total flowtime criterion
bull We have proposed four simple methods basedon local search Iterated Greedy and IteratedLocal Search
bull We have carried out a comprehensivecomparison against 12 other methods
bull IGA and ILS are state-of-the-art methods
3 Flowtime local search
httpsoaities
bull IGA and ILS better than other much morecomplex algorithms
bull pIGA and pILS also good but not better thanthe simpler non population variants
bull Future work
bull Sequence dependent setup times andflowtime objective seldom studied
bull Earliness-Tardiness objectives
bull Hybrid flowshops
3 Flowtime local search
httpsoaities
bull MIP models have been gradually abandonedand have given way to heuristics and mostnotably metaheuristics
bull However solvers keep improving Accordingto Bixby (2017 ECCO conference) since1991 solvers are now 13 million times faster(accumulated) not considering hardwareimprovements
4 Back to mathematical modeling
httpsoaities
bull Constraint Programming (CP) is gainingtraction in scheduling as of late (Bukchin andRaviv 2018 Laborie et al 2018)
bull CP models are notorious for their ability to findinteger solutions
bull However classical CP lacks boundingmechanisms and therefore does not provideoptimality guarantees or gaps
bull Since CPLEX 128 CP Optimizer guaranteesbounds and optimality GAPs for any integersolution found
4 Back to mathematical modeling
httpsoaities
bull How do modern solvers and formulationsfare for different scheduling problems
bull How does CP Optimizer compare with aregular MIP solver
bull A exact approach renaissance forscheduling problems
4 Back to mathematical modeling
httpsoaities
bull Manne (1960) MIP
4 Back to mathematical modeling
httpsoaities
bull CP Optimizer code
4 Back to mathematical modeling
httpsoaities
bull We use Taillardrsquos (1993) benchmark (120 instances) as wellas Vallada Ruiz and Framinan (2015) benchmark with 240large instances (up to 800 jobs and 60 machines)
bull CPLEX 128 + Python 366
bull Cluster of 200 virtual machines with 2 virtual processorsand 8 GBytes of RAM memory each The virtual machinesrun Windows 10 Enterprise 64 bits Virtual machines arerun in an OpenStack virtualization platform supported by12 blades each one with four 12-core AMD Opteron AbuDhabi 6344 processors running at 26 GHz and 256 GB ofRAM for a total of 576 cores and 3 TBytes of RAM
4 Back to mathematical modeling
httpsoaities
bull CP never runs out of memory and always finds asolution
No Solution Out of memoryCP 0 0
Taillard600 0 0
1800 0 03600 0 0
VRF600 0 0
1800 0 03600 0 0
MIP 642 (297) 187 (83)Taillard
600 40 01800 40 03600 34 0
VRF600 229 11
1800 166 733600 133 103
4 Back to mathematical modeling
httpsoaities
bull CP still not competitive with metaheuristics (lt03RPD)
Average GAP Average RPDCP 23168 811Taillard
600 686 3621800 609 2913600 566 253
VRF600 74924 1704
1800 41681 11793600 20543 1080
MIP 36142 1273Taillard
600 34550 8021800 31581 6113600 30532 574
VRF600 -- --
1800 41725 20843600 42322 2296
4 Back to mathematical modeling
httpsoaities
bull We have tested the following problems
bull Non-permutation flowshop with makespan criterion
bull No-wait flowshop with makespan criterion
bull Sequence-dependent setup times flowshop withmakespan criterion
bull Flowshop with Total Completion Time criterion
bull Distributed flowshop with makespan criterion
bull Hybrid flowshop with makespan criterion
bull Each problem with its own benchmark
4 Back to mathematical modeling
httpsoaities
bull Non-permutation flowshop with makespan criterionAverage GAP No Solution Out of memory
CP 22027 0 0Taillard
600 740 0 01800 663 0 03600 617 0 0
VRF600 74116 0 0
1800 39291 0 03600 16734 0 0
MIP 56102 122 422Taillard
600 55838 7 31800 53428 0 103600 52005 0 10
VRF600 59287 111 59
1800 59545 4 1683600 56510 0 172
4 Back to mathematical modeling
httpsoaities
bull No-wait flowshop with makespan criterionAverage GAP No Solution Out of memory
CP 87706 0 0Taillard
600 27031 0 01800 5610 0 03600 5543 0 0
VRF600 257211 0 0
1800 141131 0 03600 89713 0 0
MIP 167467 312 30Taillard
600 118940 0 01800 109397 0 03600 104093 0 0
VRF600 242346 107 10
1800 222293 107 103600 207732 98 10
4 Back to mathematical modeling
httpsoaities
bull Sequence-dependent setup times flowshop withmakespan criterion (480 instances)
Average GAP No Solution Out of memoryCP 5032 0 0
600 5157 0 01800 5009 0 03600 4928 0 0
MIP 124442 23 28600 131461 11 0
1800 129140 7 03600 112725 5 28
4 Back to mathematical modeling
httpsoaities
bull Flowshop with Total Completion Time criterion(360 instances)
Average GAP No Solution Out of memoryCP 55712 0 10
600 70226 0 13600 41198 0 9
MIP 29593 404 138600 30880 266 14
3600 28306 138 124
4 Back to mathematical modeling
httpsoaities
bull Distributed flowshop with makespan criterion (600instances)
Average GAP No Solution Out of memoryCP 19424 0 0
600 19743 0 01800 19162 0 03600 19367 0 0
MIP 18781 516 119600 19977 234 20
1800 16617 131 503600 19749 151 49
4 Back to mathematical modeling
httpsoaities
bull Hybrid flowshop with makespan criterion (1440instances)
Average GAP No Solution Out of memoryCP 28192 0 0
600 28491 0 01800 27954 0 03600 28130 0 0
MIP 114433 1263 0600 126879 521 0
1800 100786 314 03600 115633 428 0
4 Back to mathematical modeling
httpsoaities
bull CP is vastly superior to the best combination ofMIP model + state-of-the-art solver
bull CP gives an integer solution in all cases
bull CP does not show out of memory problems
bull Greatly reduced relative percentage deviationsand GAP values in most scenarios
bull CP still not competitive with state-of-the-artmetaheuristics but getting somewhat close insome cases
4 Back to mathematical modeling
httpsoaities
bull Companies still not using finite capacityschedulers for production programming in mostcases
bull Trend to address more realistic problems
5 Hybrid flowshops
httpsoaities
62 Maacutester-UPV
5 Hybrid flowshops
httpsoaities
63 Maacutester-UPV
5 Hybrid flowshops
httpsoaities
64 Maacutester-UPV
5 Hybrid flowshops
httpsoaities
65 Maacutester-UPV
5 Hybrid flowshops
Raw materials
preparation
Wet milling
Dry grinding Wetting Kneading Extrusion
Spray drying Molding Drying
Single fast firing
1st fast firing
Single slow firing
2nd fast firingGlazing
Glaze preparation
1st slow firing
Classification and packing
httpsoaities
66 Maacutester-UPV
5 Hybrid flowshops
Raw Materials
preparation
Milling and
grindingSpray drying Molding Drying
Glaze preparation
Single kiln firingGlazing
Selection and
packing
httpsoaities
bull We deal with hybrid flexible flowshop problemsbull A set N of jobs N=1hellipn that have to be processed on a
set M of stages M=1hellipmbull At each stage we have a set Mi=1hellipmi of mi unrelated
parallel machinesbull All jobs are processed sequentially on the stagesbull pilj is the processing time of job j at machine l inside stage i
5 Hybrid flowshops
httpsoaities
5 Hybrid flowshops
1
2
l
1
2
3
3
4
4
C11 = 4
5
1
2
1
i
2
2
1
0 1 2 3 4 5 6 7 8 9 10 11 12
C12 = 3
C14 = 6
C13 = 9
C15 = 7
5
C22 = 7
C24 = 8
C25 = 12
C23 = 10C21 = 9
httpsoaities
bull We consider many realistic constraintsbull Fj denotes the set of stages that job j visitsbull Eij is the set of machines that can process job j at stage ibull rmil represents the release date of machine l inside stage
i No job can start processing on this machine before thisrelease date
bull Pj is the set of jobs that should precede job j The firstoperation of job j should wait until all last operations ofjobs in Pj have finished
5 Hybrid flowshops
httpsoaities
bull Siljk is the machine based setup times on machine l atstage i when processing job k after having processed jobj
bull Setups might be either anticipatory or non-anticipatorybull Ailjk=0 -gt the setup is non-anticipatory and one must wait for
job j to arrive at machine l before carrying out the setupbull Ailjk=1 -gt the setup can be done as soon as machine l is free
5 Hybrid flowshops
httpsoaities
bull lagilj is the time lag between the end of job j at machine linside stage i and the beginning of the next stage in whichjob j is processed
bull If lagiljlt0 then overlapping occurs (not all parts in job j needto be finished at stage i before processing at the next stagestarts)
bull If lagiljgt0 then we have a waiting time between stages
5 Hybrid flowshops
httpsoaities
bull We will consider the common makespan criterioninitially
bull The problem is NP-Hard since many simplificationsof it are NP-Hard
bull Following the common three field notation andVignier et al (1999) the problem can be denotedas
max)(
1)( ))(( CprecMSlagrmRMHFFLm jijk
mi
i=
5 Hybrid flowshops
httpsoaities
1
2
1
2
1
1
2
3
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75
Rm11
4
1
14
Rm12
3
Rm21
Rm22
Rm31
8
16
23
2
12
2
21lag122
24lag111
1
3214
1
28 34lag221
S2123
27
li
3
36
S3113
40
3
48
S1224
42 48
S1115
44 48
4
58
4
6059 6658
S3134
lag124
5
59
5
lag11554
S2215
48 66
S3145
72 74
5lag225
77
10
59
lag21336 39
Cmax=77
5 Hybrid flowshops
httpsoaities
Job 2 Time
Machine 1
Machine 2
Machine 3
Machine 4
Stage 1
Stage 3
Job 3Job 1
Stage 2
Machine 5
Setup Release dates
Cmax
5 Hybrid flowshops
httpsoaities
t5ms25ms125ms
3
6
9
12
SRIS IG ILS MA GA
5 Hybrid flowshops
httpsoaities
bull We extend the problem with theconsideration of due dates
bull Cj is the completion time of job j at the shop
bull dj is the due date of job j
bull wj its importance or priority
bull Tj=max0 Cj - dj is the tardiness of job j
bull Ej=max0 dj - Cj is the earliness of job j
5 Hybrid flowshops
httpsoaities
bull Extension of the due date concept to due datewindows
5 Hybrid flowshops
djCj
0
dj- +
httpsoaities
bull Considered objective function
5 Hybrid flowshops
httpsoaities
bull Literature on hybrid flowshop very rich
bull This objective has not been studied yet
bull Objectives of the research
bull To propose simple methods
bull To compare them against the best adaptedalgorithms from the literature
bull State-of-the-art results
5 Hybrid flowshops
httpsoaities
5 Hybrid flowshops
bull We propose an Iterated Local searchprocedure
bull Very simple
httpsoaities
5 Hybrid flowshops
bull Solution representation is a complex issue inHFS
bull A simple permutation not enough
bull Complex representations too time consuming
bull Simple permutation with the FAM rule
bull For stage 2hellipm jobs are sorted according totheir completion times in the previous stage
httpsoaities
5 Hybrid flowshops
bull Active schedules not enough to optimizeearliness
bull We apply an idle time insertion method to alljobs in the last machine
1 2 3 4
62 64 67 68 70 72 74 76 79 81 88 90 93 95 97 99
d1 d6d1 + d2 d2 + - d3 d3 - + d4 d4 - + d5 d5 - + d6 d7 d7 d8 d8 - + - + - +
5 6 7 856 61 65 69 75 82 84 86 92 9793
1 2 3 4
62 64 67 68 70 72 74 76 79 81 88 90 93 95 97 99
d1 d6d1 + d2 d2 + - d3 d3 - + d4 d4 - + d5 d5 - + d6 d7 d7 d8 d8 - + - + - +
5 6 7 857 62 66 70 76 83 86 88 9894
httpsoaities
5 Hybrid flowshops
bull We apply a two stage local search
bull The first one works with the permutation(insertion and interchange inside VND)
bull A second local search is carried out over theexact representation
bull But only a few promising neighbors areexplored
httpsoaities
5 Hybrid flowshops
bull Simple perturbation by carrying out anumber of insertion and interchanges
bull Directed perturbation A number of perturbedsolutions is generated and the best isobtained
bull Tournament acceptance criterionbull List with the best found solutions
bull If a better solution is found the list is cleared
bull If solution not better tournament select a solution fromthe list