1 the modelling potential of minimax algebra h.p. williams london school of economics
TRANSCRIPT
11
THE MODELLING POTENTIAL THE MODELLING POTENTIAL OF MINIMAX ALGEBRAOF MINIMAX ALGEBRA
H.P. WilliamsH.P. Williams
London School of EconomicsLondon School of Economics
22
MaxMax
MinMin
++
-- // //
N.B. (unlike +) does not have an inverse. Hence we are concerned with semi rings.
33
Connections between Minimax Algebra and Connections between Minimax Algebra and Mathematical ProgrammingMathematical Programming
1. Mathematical Programming is concerned with models of form
)""unlike(i.e.
dbca
dbca
dc
ba
Duality Theorem of LP demonstrates inequalities can be added in multiples to give tight bound on optimal objective value
ixgxf
i 0)(:tosubject)(Max/Min
2. Can add but not subtract inequalities
44
The Dual of a Disjunctive Programme (in The Dual of a Disjunctive Programme (in Disjunctive Normal FormDisjunctive Normal Form))
0j
jj
x
bxa
xc
kijkijik
j
Minimise
Subject to:
Dual is
Maximise kii
kik
ybSubject to:
0ki
k ikij
y
ca
If Primal and Dual Solvable (not Infeasible or Unbounded) they have the same optimal objective value.
ik ,
j
ik ,
'
j j
55
Dynamic ProgrammingDynamic ProgrammingSpecial cases lend themselves to Mini Max formulations
Knapsack Problem
11
1
1
,1,Max),1(
:Subject to
Maximise),(
nnn
j
n
jj
j
n
jj
cabnfbfbnf
bxa
xcbnf
(a nested recursion)
Can be written as
),(//,1,1 11 bnfabnfcbnf nn
66
The Group Knapsack ProblemThe Group Knapsack Problem
1. Solve LP Relaxation of an Integer Programme
2. Choose integral, non-negative values for non-basic variables so as to make basic variables integral
3. This gives a set of congruence relations i.e. a group equation (in non-negative variables)
i.e. gxgojj
n
ij
Where are members of a finite abelian group gj
77
The Group Knapsack ProblemThe Group Knapsack Problem
4. If we choose so as to minimise difference from optimal LP objective we have an objective
Minimise Where are LP reduced costs of non-basic
variables
jx
j
n
j
jxc1
jc
As with conventional Knapsack Problem can be solved by Dynamic Programming
)()(( ,//,11 1),1 bnagnn ffnf ncg
88
The Group Knapsack ProblemThe Group Knapsack Problem
5. But resultant solution may imply negative (infeasible) values for basic variables.
Need therefore to seek 2nd best, 3rd best etc. solution to Group Knapsack Problem
6. Cuninghame-Green does this systematically (‘Integer Programming by Long Division’) by successively enumerating solutions of monotonically increasing cost until a feasible solutions if found.
99
Shortest Path ProblemShortest Path ProblemAll deterministic Dynamic Programmes (including Knapsack problems) can be formulated as Shortest Path problems
0
.
.
..
0
)( arc oflength is
(t) node terminalto node frompath shortest is
dd
ei
if
ifdaMindijd
id
jij
j
i
i
i
Different Linear Algebra methods mirror different Shortest Path methods
trow
tj tj
1010
Minimax ProblemsMinimax Problems
Facility Location, Obnoxious Facilities, Political Districting, Nucleolus of a Game, ‘Fair’ Allocations.
ii
i xMinxMaxx
Min
Analogous to conventional objectives
1111
LL1 1 and L Normsand L Norms
Can formulate Linear Regression problems using Minimax Algebra
Special Case: Zero Dimension problems
Median of a Set of Numbers
)(,...,, 121 NormLyyy n
Pi Pi
Pii yyMaxNNorm is a partition of numbers into sets of
PP :
)(2
1
)(2
oddnnumbersn
or
evennnumbersn
i.e. Maximum ‘Contrast’
Pii
Pi
i
y
yN
1212
L NormL Norm
j
i
ji
jiji
y
yN
yyMaxNNorm
,
,2
1
1313
Scheduling ProblemsScheduling Problems
Jobs cannot start (e.g. Until previous jobs (e.g. ) have finished (durations )
3x 21, xx21,dd
22113
22113
..,
xdxdxeidxdxMaxx
May wish to Minimise completion of all jobs
iii
i
xdMinei
xdMaxMinge
..
.. 11
Mirrors an LP model
x
1414
ReferencesReferences
BA CarrBA Carré, An Algebra for Network é, An Algebra for Network Routing Problems, Routing Problems, J Inst Math Appl J Inst Math Appl 7 7 (1971) 273-293(1971) 273-293
RA Cunningham Green RA Cunningham Green Minimax Algebra Minimax Algebra Vol. 166. Lecture Notes in Economics and Vol. 166. Lecture Notes in Economics and Mathematical Systems, Springer Verlag Mathematical Systems, Springer Verlag 19791979