579
1
INTRODUCTION
Manypracticalproblemsgiverisetosystemsoflinear
inequalitiesas
*
Ax b (1)
inwhichtheinequalitiesarecomponentwise,i.e.

n
*
ij j i
j1
Ax b, i 1,...,m (2)
where
ij
A is the (i, j) element of
mxn
A
and



m
i
1im
bb
. Also, let
T
A ,
A denote the
transposeandtheMoorePenrosepseudoinverseofA,
respectively.

,
, willstandfortheEuclidean
scalarproductandnorm.
If(1)isconsistent,manyclassesofefficientsolvers,
mostly iterative ones, have been designed for its
numerical solution (see for a good overview the
monograph[3]).Intheinconsistentcaseof(1),forany
*n
x itexistsatleastoneindex

i 1,...,m such
that the ith inequality in (2) is violated, i.e. the set
I x 1,...,m
,definedby

ii
I x i 1,...,m , A ,x b  (3)
is nonempty for any
n
x . Moreover, let us
supposethattheset

1p
I x i ,...,i isorderedsuch
that

12 p
ii i, then,
 
Ix Ix
A,b will denote the
submatrix of A with the rows
1p
ii
A ,...,A and the
subvector of b with components
1p
ii
b ,...,b ,
respectively. For any vector
m
y we define
m
y by

i
i
y max y ,0 , and the convex sets
by

n
iii
Cx ,A,xb,i 1,...,m . The
Regularized Han-type Algorithms for Inconsistent
Maritime Container Transportation Problems
D.Carp
ConstantaMaritimeUniversity,Romania
C.Popa&C.Şerban
OvidiusUniversityofConstanta,Romania
ABSTRACT:InthispaperweanalyseseveralwaystocomputetheweightsfromtheRegularizedHan(RH)
algorithm, the regularized version of Han’s algorithm for approximating the least squares solutions of
inconsistent (incompatible) systems of linear inequalities. We tested our approaches on a classical
transportation problem, aiming to provide a cost optimized solution to real world transportation problems,
whichoftenareunbalancedandinconsistent.
http://www.transnav.eu
the International Journal
on Marine Navigation
and Safety of Sea Transportation
Volume 8
Number 4
December 2014
DOI:10.12716/1001.08.04.13
580
inconsistency of the system (1) is equivalent to
m
i
i1
C0,andwereformulateitina“leastsquares
sense”as(seee.g.[5]):find
*n
x suchthat




*n
fx minfx,x  (4)
with
 

2
1
fx Ax b
2
Thefollowingresultsareprovedalsoin[5].
Proposition1.
(i) The objective function f from (4) is continuously
differentiable,convexand(seealso(3))

  


TT
Ix Ix Ix
fx AAxb A A xb (5)
(ii) Thereexists(atleast)aleastsquaressolutionof
(1).
(iii) Avector
*n
x of(1)ifandonlyif


*T*
fx A Ax b 0 (6)
The paper is organized as follows: in sections 2
and 3 we present Han’s original algorithm for
inconsistent systems of linear inequalities and the
Regularized version of it, respectively. Section 4
presents the characteristics of a classical
transportation problem and the way it can be
modelled so that the
Hantype algorithms can be
appliedon.Section5isdedicatedtothemethodswe
usedforchoosingthebestweightsoftheRegularized
Hanalgorithm,thenumericalexperimentsbeingdone
over an unbalanced and inconsistent transportation
problem.
2
HANTYPEALGORITHMFORINCONSISTENT
SYSTEMSOFLINEARINEQUALITIES
S.P. Han proposed in [5] the following iterative
algorithmforapproximatingaleastsquaressolution
ofthesystem(1):
AlgorithmH.
Let
0n
x beaninitialdatum;for k 0,1,... do:
Step 1. Find

k
k
IIx and compute
kn
LS
d as
the (unique) minimal norm solution of the linear
equalitiesleastsquaresproblem


kkk
k
III
Ad b Ax min! (7)
Step 2. Compute
k
as the smallest minimizer
ofthefunction


kk
LS
fx d , (8)
Step3.Set

k1 k k k
LS
xx d
Theexistenceofthesmallerminimizerfortheconvex
functionfrom(8)wasprovedin[5]andaprocedure
tofinditwasgivenin[1].
Theorem1.
(i) Let
k
k0
x be the sequence generated by
algorithm
Hfromany
0n
x .Then,

k
fx 0,for
a
k or

k
k
limf x 0 .
(ii) For any m x n matrix A, any righthand side
m
b and any initial datum
0n
x , Han’s
algorithm
Hproducesaleastsquaressolutionofthe
system (1) in a finite number of steps (in exact
arithmetic).
3
REGULARIZEDHAN(RH)ALGORITHM
The Regularized Han algorithmwas introduced in
[10]andthoroughlystudiedin[9].In[10]theauthor
comments on Han’s algorithm by considering its
major drawback in the fact that, in each iteration,
initialobjectivefunctionfrom(4)isreplacedby
 

kk
2
2
kk k
II
11
fx Ax b Ax b
22
.
In this way, many originally satisfied constraints
might be violated in the new iterative solution
k
x
.
Regarding this aspect, he also proposed to use the
complement of the set
k
I , denoted by
k
J and
characterizedby

kk
kii
J J x i 1,...,m ,A x b  (9)
together with a diagonal weights matrix
k
kkk
k1qi
W diag w ,...,w ,w 0 , where
k
q is the
numberofelementsintheset

k
J, k 0
.Withthese
ideas he designed the Regularized version of Han’s
algorithmfrombelow.
AlgorithmRH.
Let
0n
x
beaninitialdatum;for
k 0,1,...
do:
Step 1. Find
k
k
IIx,

k
k
JJx as before and
compute
kn
d the minimal norm solution of the
(regularized)linearleastsquaresproblem


kkk k
2
2
k
III kJ
Ad b Ax WAd min!
 (10)
Step2.Compute
k
asthesmallestminimizerof
thefunction
581




kk
fx d , (11)
Step3.Set

k1 k k k
xx d.
Theorem2.
Let
0n
x be arbitrary fixed, and

k
k0
x the
sequencegeneratedbythealgorithm
RH.Then,either
itexists
0
k0 such that

0
k
fx 0,or


k
k
limf x 0 .
4
THETRANSPORTATIONPROBLEM
Aclassicaltransportationproblemimpliesfindingthe
minimumcostof transportingcertainquantitiesofa
single type of commodity from a given number of
loading ports (sources) to a given number of
unloadingports(destinations).
At each source


i
i 1,...,n
S , supplies

i
i
s 1,...,n
ofsomegoodsareavailable,andateachdestination


j
j
1,...,m
D some demands

j
j
d1,...,m
are
requested. Table 1 illustrates the classical
transportation problem,


ij
i 1,...,n ,j 1,...,m
c being the
costsofshippingoneunitofcommodityfromsource
i
S todestination
j
D .
Table1.Theclassicaltransportationproblem
_______________________________________________
D1 D2 D3  DmSupply(s)
_______________________________________________
S1c11 c12 c13  c1ms1
S2c21 c22 c23  c2ms2
      
S
ncn1 cn2 cn3  cnms n
_______________________________________________
Demand(d) d1 d2 d3  dm
_______________________________________________
If we denote by 
ij
x,i 1,...,n,j 1,...,m the
number of units transported from source
i
S to
destination
j
D , we get the following mathematical
modelofthe(classical)transportationproblem:


nm
ij ij
i1j1
min c x (12)

n
ij j
i1
s.t. x d ,j 1,...,m (*)

m
ij i
j1
xs,i 1,...,n (**)

ij
x0,i 1,...,n,j 1,...,m. 
Remark1.
Some arguments for the relations (*)(**) are as
follows:
for(*):ateachdestination,thedemandhasto be
”at least” satisfied (e.g. the construction of a
buildingwillnotbestarted if we do not have at
leastaminimalamountofmaterials)
for(**):allavailableunitsmustbesupplied.
Theproblemiscalled
balancedifthetotalsupply
equals the total demand (i.e.


nm
ij
i1 j1
sd) and
unbalanced otherwise. In the balanced case or the
unbalancedonewith

nm
ij
i1 j1
sd,thelinearprogram
(12)isconsistentandwellknownmethods(including
Simplextype algorithms) are available (see [4]). We
willconsiderinthispapertheunbalancedcase

nm
ij
i1 j1
sd (13)
for which the linear program (12) becomes
inconsistent.
Let us suppose that there exist 7 sources of
containers
17
S ,...,S and7warehouses
17
D ,...,D .
Wewillconsidertheunbalancedandinconsistent
transportationproblem
PdescribedinTable2.
Table2.Theunbalancedtransportationproblem(P)
_______________________________________________
D1 D2 D3 D4 D5 D6 D7 Supply(s)
_______________________________________________
S13 3 4 12 20 5 9 1050
S
27 1 5 3 6 8 4 350
S
35 4 7 6 5 12 3 470
S
44 5 14 10 9 8 7 600
S
58 2 12 9 8 4 2 600
S
66 1 8 7 2 3 1 480
S
79 10 6 8 7 6 5 450
_______________________________________________
Demand(d) 455 320 540 460 760 830 780
_______________________________________________
Afterapplyingaseriesofrefinements(see[2]),we
canwritetheproblem
Pasthelinearprogram
min c,y s.t.By d,y0 (14)
withthecorrespondingdualproblemgivenby
T
max d,u s.t.Bu c,u0 (15)
In [2] we also proved a result that gives us the
possibilitytoexpressinanequivalentwaytheprimal
dual pair of linear programs (1415) as the linear
systemofinequalities(1)where









TT
T
0
cd
d
B0
A,b
c
0B
0
I0
0
0I
(16)
So,insteadofsolvingthelinearprograms(14‐15),
onecouldsolvethesystem(1)byapplyingtheHan
typealgorithms.
582
5
COMPUTINGTHEREGULARIZATION
PARAMETERS
The success of Regularized Han algorithm depends
on making a good choice of the weights


k
kkkk
1qi
w w ,...,w ,w0. In the literature,
k
w are
called the regularization parameters. Hence, one
question that can occur related to Regularized Han
algorithmis:howcanwecomputetheregularization
parameters
k
w so that the RH algorithm preserves
itspropertiesgivenbyTheorem2?In[10],theauthor
only remarks that we can impose the appropriate
penalties when
k
x approaches


kk
j
jj
Hj,Axb
byassigningalargerweight
k
w forjthequationif
thecurrentiterativesolutionisclosetotheboundary
j
H ,andasmallerweightifisfaraway.Takingthese
remarksintoconsideration,weanalysedseveralways
forchoosinggoodregularizationparameters,ourgoal
beingtocomputeagoodestimateofthesolutionto1.
Wetestedthemethodsontheinconsistentproblem
P
describedinsection4,knowingthatforthisproblem,
theminimalcostsolutionis15336,obtainedwithHan
algorithm. Both algorithms,
H and RH, were
implemented in Matlab R2010a, using the builtin
Matlabfunctionpinvtocomputethedirection(Step1
of the algorithms), all runs being started with the
initial datum

T
T
00
xy,0 with
0
y0
and being
terminated if at the current iterations
k
x satisfy

Tk
AAx b AT(Axk−b)+i.
Thefirstmethodconsidered(
M1)triestoseteach
k
j
k
w,j 1,...,q according to
kk
k
JJ
Ax b . We can
implementitunderthe
M1
fork=0,1,...do


kk
kk
J
J
hAxb
forj=1,…,
k
q
do
if
k3
h10
then
k
j
k
1
w
h
else
k3
j
w10
Table3.Thecostsolutionofproblem(P)
_______________________________________________
HanRHwithM1
_______________________________________________
1533615981
_______________________________________________
Next,wetooksomefixedweightsateachstepk,
bysettingeach
k
j
k
w,j 1,...,q equalto
kk
k
JJ
Ax b
M2
fork=0,1,...do


kk
k
j
k
k
JJ
1
w , j 1,...,q
Ax b

Table4.Thecostsolutionofproblem(P)
_______________________________________________
HanRHwithM1
_______________________________________________
1533616257
_______________________________________________
The third method considered uses fixed weights
forallstepsk
M3

k3
j
k
w 10 , k, j 1,...,q
Table5.Thecostsolutionofproblem(P)
_______________________________________________
HanRHwithM1
_______________________________________________
1533616113
_______________________________________________
Analysing (10), we observe that we can make an
analogybetweentheRegularizedHanalgorithmand
TikhonovRegularization.

kk
2
2
II k
Ad r d min! (17)
with

kk k
k
III
rbAx and 
k
kkJ
WA . The most
popular method used to find the regularization
parameter of a Tikhonov Regularization is the L
curveapproach.AnLcurveisdefinedby
log Lx ,log Ax b , 0
and the Lcurve method selects the regularization
parameterasthecorner ofLcurve,i.e. thepointof
maximumcurvature(seefordetails[6]).
For implementation, we used Hansen’s
Regularization Tools package ([7]) to compute the
cornerofLcurvefor(10).
M4
fork=0,1,...do
lcorner=l_curve(U,S,rz’,Tikh’,L,V)


k
j
k
wlcorner, j 1,...,q
583
where
1
[U,S,V] = csvd(
k
I
A ) where ‘csvd’ stands for
CompactSVD:
k
T
Irrr
AUSV
2

kk
k
II
rz b A x
3
k
J
LIA
Table6.Thecostsolutionofproblem(P)
_______________________________________________
HanRHwithM1
_______________________________________________
1533624703
_______________________________________________
AfterimplementingM4onproblemP,wenoticed
thatanadjustmentisneedtobemade.
M5
fork=0,1,...do
lcorner=l_curve(U,S,rz’,Tikh’,L,V)


k
j
k
wlcorner, j 1,...,q



33
iflcorner 10 thenlcorner 10


k
j
k
wlcorner, j 1,...,q
Table7.Thecostsolutionofproblem(P)
_______________________________________________
HanRHwithM1
_______________________________________________
1533615336
_______________________________________________
REFERENCES
[1]Carp,D.,Popa,C.&SerbanC.2013.Iterativesolutionof
inconsistent systems of linear inequalities, Proceedings
on Applied Mathematics and Mechanics (PAMM), 13,
DOI10.1002/pamm.201310199,407–408.
[2]Carp, D., Popa, C. & Serban C. 2014. Modifed Han
algorithm for maritime containers transportation
problem, ROMAI Journal, v.10, no.1, ISSN 1841
5512,
ISSNOnline20657714,11–23
[3]Censor, Y. & Stavros, A.Z. 1997. Parallel optimization:
theory,algorithmsandapplications, Numer.Math.and
Sci.Comp.Series,OxfordUniv.Press,NewYork
[4]Koopmans, T.C. & Beckmann, M. 1957. Assignment
problems and location of economic activities,
Econometrica,Vol.25,No.1,53–76.
[5]
Han, S.‐P. 1980. Least squares solution of linear
inequalities,Tech.Rep.TR2141,MathematicsResearch
Center,UniversityofWisconsin‐Madison.
[6]Hansen, P. C. & O’LearyD. P. 1993. Theuse oftheL
curve in the regularization of discrete illposed
problems,SIAMJ.Sci.Comput.,14,14871503.
[7]Hansen, P. C. 1992. Regularization Tools, a Matlab
PackageforAnalysisandSolutionofDiscreteIllposed
Problems, Technical Report UNIC9203, Danish
Computing Center for Research and Education,
TechnicalUniversityofDenmark,(RevisedJuly2007).
[8]Popa,C.1998.Extensionsofblock‐projectionsmethods
with relaxation parameters
to inconsistent and rank
defficientleastsquaresproblems,BIT,38(1),151–176.
[9]Popa, C. & Serban, C. 2014. Han‐ type algorithms for
inconsistent systems of linear inequalities‐a unifed
approach, Applied Mathematics and Computation,
Volume 246, ISSN: 00963003, DOI:10.1016/
j.amc.2014.08.018,247–256
[10]Yang, K. 1990. New iterative methods
for linear
inequalities,Tech.report906,DepartmentofIndustrial
and Operations Engineering, University of Michigan,
USA