evolution-ary algorithm
Forthemeta-heuristi algorithmwerstneedtoadjustthemetaheuristi parameters
for the algorithms. This is in many asesmore of an art, than a s ien e. In the
following we will experimentally ndthe best parametersetting. The test asewe
use to adjust the parametersfrom, is the HAL omputation with a time frame of
T = 20
. This is an arbitrary ase, and there is no guarantee this will leadto the−1 0 1 2 3 4 5 6 7
Figure 7.4: Solution (HAL
T = 20
) from simulated annealing asa fun tion of theα
temperature hange oe ientand thenumberN
of iterationsto rea hthermalequilibrium.
Figure7.5: Solution(HAL
T = 20
)fromevolutionaryalgorithmasafun tion oftheG stop
generation ountandthepopulationsize (N
).optimal set of parameters for all other ases. In parti ular one should beware of
ne-tuningthe algorithmpre isely tothis aseasit mightmeanthemeta-heuristi
algorithmsis reallygoodatndingthissolution,butterribleforallother asesand
ALU
Figure7.6: S hedule,FUallo ationandoperatorassignmentgeneratedbysimulated
annealingforHALwith
T = 20
onstraint,givingatotalbalsa- ost areaof59700
.smallexamplewheretotheexa toptimumisknownisgoodtestfornarrowingdown
theparametersetting.
Webeginwithsimulatedannealing,whereweneedtondthetemperature hange
oe ient
α
andthethermalequilibrium numberN
. InFigure7.4wehaveshownseveralrunsofthealgorithmforvariousparametersettingsandplottedthesolution
thealgorithmnds. Ea hpointrepresentsanentirelynewrun. As anbeseenthe
simulatedalgorithm isratherunstable apableofgettingstu katalo alminimum.
Howeverfor
N = 500
and larger, the algorithm tends to be ome more stable andprodu e good solutions(a tually the optimalsolution) at everyrun. The best
pa-rameter setting for
α
seems to beα = 1.250
for largerα
the algorithm does notprodu e better solutions, only taking exponentiallymore time to omplete. These
settingalsoseemtoprodu egoodsolutionsfortheotherproblemsintheben hmark
set.
Nextis the evolutionaryalgorithm, whereweneed to ndthe
G stop
generationount and the
N
population ount. In Figure 7.5 we have shown several runs ofthealgorithmforvarious parametersettings andplotted thesolutionthealgorithm
nds. Againea hpointrepresentsanentirelynewrun. As anbeseenthesimulated
algorithm is rather stable apable of produ ing reliable results. Another fa tor is
the high-dependen y on the population size. With a population around
512
thealgorithm starts onverging towardsthe global optimumwith the fast onvergen e
and hoosing a large population size doesnot in rease the onvergen e. The best
value for the maximum generation ount
G stop
seems to be around in the rangefrom
320
to640
. To be on the safe side we hose640
generations. Again these parameterssettingsseemstoprodu egoodsolutionsfortheotheralgorithmsin thesomeparti ularsolutions.
We have ben hmarked the algorithms on two DFGs: HAL (
T ASAP = 10
) andCOSINE(
T ASAP = 11
). WeareinterestedintheCPU-timei.e.. theamountoftimeittakesrunningthealgorithmstogetasolutionsatisfyingourarearequirements. For
the twoDFGs weapply thetwometa-heuristi algorithms, givingus four primary
test ases(shownintable7.4). Forea htest asewesetvesili onarearequirements
and sixtime framerequirements
T = dt + T ASAP
,(the blanks arewhere themeta-heuristi -algorithmsfailtondasolutioneitherbe ausethereisnooptimalsolution
satisfyingtherequirementorin border asesbe ausethealgorithmsareheuristi ).
Again,alltestsareperformedona200MHzPentiumII,with96MBmemoryand
allnumbersree tastatisti alaverageofrunningthealgorithms500timesonea h
problem instan e.
In general the simulated annealing out-performs the evolutionary algorithm in
termsofCPUtimerequiredtondasolutionforlargeproblems(i.e. COSINE).The
primary reasonstems from the evolutionary algorithm working on a large
popula-tion,whi hineveryiterationhastobemadefeasibleand ostevaluated,whereasthe
simulatedannealing onlyworks withoneproblem instan e. Ontheother-handthe
evolutionaryalgorithm seems to performmore stable,unlikesimulated annealing
whi h is apable of getting stu k in lo al-minimums for some runs. Comparing
theevolutionaryalgorithmwiththesimulatedannealingtheevolutionaryalgorithm
takes signi antlylonger time torun anddoes ndjust asgood solutionsas
simu-lated annealing. Inparti ular in the COSINE ase theevolutionaryalgorithm has
problems. Thisdoesnotmeantheevolutionaryalgorithm annot ndthesolutions
eg. if runfree the evolutionary algorithm is apable of nding a solution for
CO-SINE,
T = 107
, belowthe arearequirementof49200
, howeverit took25857.4s
orapproximately7.18hours. Theevolutionaryalgorithmdoesnothoweverhavesimilar
problemsforFIR orELLIPTIC.
A property of the proposed CDFG synthesis algorithm is that one of these
al-gorithms will be run for the DFG fragments, until the main synthesis algorithm
onverges,itisthereforeimportantthatthesealgorithmsgeneratethesolutionsfast.
This favoursthesimulatedannealingoverthetwootheralgorithms.
Finally in Figure 7.6 is shown the optimal s hedule generated by the
meta-heuristi algorithmsin theparameterinvestigation.