Академический Документы
Профессиональный Документы
Культура Документы
: The index set of (Y, W ) that is defined as J = 1, 2, ..., |(Y, W )| where |(Y, W )| presents
the cardinality of (Y, W ).
(EC.1.1)
zk
Ak x +
eks wks + Ck zk bk
kK
(EC.1.2)
z k Zk
kK
(EC.1.3)
kK
0
gk (x, eks ) = min
c|2k zks
0
(EC.1.4)
zks
0
Ak x + eks + Ck zks
bk
kK
(EC.1.5)
0
zks
Zk
kK
(EC.1.6)
For each adversary scenario (y, w) (Y, W ) with index j J, with respect to constraints
j
s Sk is equal to 1 for each k K. Let sj
(17) and (25), exactly one of the variables wks
j
denote the index in Sk for which wks
is equal to 1. Therefore we have the following relations.
j
j
wks
=1
j
jJ
(EC.1.7)
j
wks
=0
j J, s 6= sj
(EC.1.8)
In the following we prove the if -statement of Lemma 1. The only if -statement of this
lemma can be proven in a reverse direction. Assume that x
is a first-stage feasible solution
of Model (P1). In the following we separately prove that
ec2
- x
is also a first-stage feasible solution of Model (P2).
- The objective values of (P1) and (P2) for this fixed first-stage solution are the
same if the adversary and the decision maker choose their optimal solutions in the
adversarys problem and the second stage subsequently.
Proof of Part 1: Since Model (P1) is feasible, there is at least a feasible second-stage policy
hkj i(kK) for each j J such that
Ak x
+
j
+ Ck kj bk
eks wks
k K, j J
(EC.1.9)
kj Zk
k K, j J
(EC.1.10)
kK
k K, j J
(EC.1.11)
kj Zk
k K, j J
(EC.1.12)
Therefore, x
is also a first-stage feasible solution of Model (P2).
Proof of Part 2: To prove that the objective values of (P1) and (P2) for the fixed first-stage
solution x
are the same, it is enough to prove that relation (EC.1.13) or its equivalent, relation
(EC.1.14), holds.
P
|
max
c1 x
+
fk (
x, w) =
(y,w)(Y,W )
kK
max
(y,w)(Y,W )
kK
fk (
x, w) =
!
max
(y,w)(Y,W )
c|1 x
+
P P
gk (
x, eks )wks
(EC.1.13)
kK sSk
!
max
P P
(y,w)(Y,W )
kK sSk
gk (
x, eks )wks
(EC.1.14)
gk (
x, eksj ) and fk (
x, wj ) have the same structure and following relations hold.
gk (
x, eksj ) = fk (
x, wj )
k K, j J
(EC.1.15)
arg min gk (
x, eksj ) = arg min (fk (
x, wj ))
k K, j J
(EC.1.16)
0
zks
zk
The following stream of equalities proves the validity of (EC.1.14). In the following relations the second equality is obtained using (EC.1.15). The third equality is valid because of
(EC.1.7)-(EC.1.8).
ec3
P
P
j
gk (
x, eksj ) =
fk (
x, w ) = max
max
fk (
x, w) = max
jJ
jJ
(y,w)(Y,W ) kK
kK
! kK
!
P
P
P P
j
= max
gk (
x, eks )wks
max
gk (
x, eks )wks
jJ
kK sSk
(y,w)(Y,W )
kK sSk
In addition, (EC.1.18) shows that we can obtain the second-stage optimal policies for
0
.
variables zk in Model (P1) from the optimal values of variables zks
ec4
kK sSk
where gk (
x, eks ) is defined as follows.
0
gk (x, eks ) = min
c|2k zks
0
(EC.2.2)
zks
0
bk
Ak x + eks + Ck zks
kK
(EC.2.3)
0
Zk
zks
kK
(EC.2.4)
0
for k K, s Sk are independent of (y, w)
It is clear that the optimal values of vectors zks
k K, s Sk
(EC.2.5)
0
0
where Gks = {zks
Zk |Ak x
+ eks + Ck zks
bk }. Therefore, because of the independence of
0
zks
, k K, s Sk from (y, w) (Y, W ), we can swap max and min
in Model (P2) and Theorem
0
(y,w)
2 is proven.
(z )
ec5
min
max
(x,z 0 )(X,Z 0 )
(y,w)(Y,W )0
c|1 x +
0
wks
c|2k zks
P P
(EC.3.1)
kK sSk
where (Y, W )0 = {(y j , wj ), j = 1, 2, ..., m}. Since (Y, W )0 (Y, W ) the optimal objective value
of Model (MP0 ) is a valid lower bound for the optimal objective value of the original robust
problem (P4). In the following we demonstrate that (MP0 ) is equivalent to (MP). By writing
the convex combination of m scenarios (y j , wj ), Model (MP0 ) can be rewritten as follows.
(MP00 )
min
(x,z 0 )(X,Z 0 )
max c|1 x +
m
P
j=1
m
P
!!!
P P
j
0
w
ks
c|2k zks
(EC.3.2)
kK sSk
j = 1
(EC.3.3)
j=1
j 0
j = 1, 2, ..., m.
(EC.3.4)
In Model (MP00 ), for a fixed value of (x, z 0 ), the inner max problem is a linear programming
model and one of its extreme points will be the optimal solution. Each extreme point of this
model corresponds to one of the scenarios (y j , wj ). Therefore, Model (MP00 ) is equivalent to
Model(MP0 ). By dualizing the inner max problem in Model (MP00 ) and assuming as the
dual variables of constraint (EC.3.3) we obtain Model (MP) and Theorem 3 is proven.
ec6
XX
0
w
ks Opt
c|2k zks
(EC.4.1)
kK sSk
where is the optimal solution of the master problem in this iteration and Opt is the optimal
objective value of the robust problem. Besides, in the Benders algorithm without stopping
conditions for the master problem and subproblem, the adversary scenario (
y , w)
is visited in
the subproblem if it is the optimal solution of the subproblem. Therefore, we have
Opt c|1 x
+
XX
0
c|2k zks
w
ks
(EC.4.2)
kK sSk
Relation (EC.4.2) is valid because the optimal objective value of the subproblem is a valid
upper bound for the optimal objective value of the robust problem. Consequently, (EC.4.1)(EC.4.2) results in relation (EC.4.3).
Opt = c|1 x
+
XX
0
c|2k zks
w
ks
(EC.4.3)
kK sSk
ec7
ec8
XX
0
w
ks
c|2k zks
(EC.5.1)
kK sSk
c|1 x
+
XX
0
c|2k zks
w
ks +
(EC.5.2)
kK sSk
Obviously relation (EC.5.2) is in contrast with (EC.5.1) and we conclude that if the algorithm visits an adversary scenario with a repeated vector w in the subproblem, this adversary
is the optimal solution of the subproblem. To prove that the optimal objective value of the
subproblem is equal to the upper bound of the master problem in iteration i 1, we have to
show that in the master problem, an instance of constraint (42) corresponding to the repeated
vector w
is binding. If for another adversary scenario with a different repeated vector w = w0 ,
P P | 0 0
P P | 0
c2k zks w
ks < c|1 x
+
c2k zks wks
constraint (42) is binding, then we must have c|1 x
+
kK sSk
kK sSk
ec9
holds. In iteration i, since the algorithm found an adversary scenario with a repeated vector
w in subproblem SP(t), regarding Lemma 2 this adversary scenario is the optimal solution of
= OiSP holds. Further, in the master problem M P (i) that is solved
the subproblem and LSP
i
after subproblem SP (i), two cases are possible.
Case 1) OiM P > OiSP holds. First note that OiM P < Opt is a valid regarding Theorem
3. OiM P > OiSP together with OiM P < Opt results in Opt > OiSP . The latter relation
contradicts with the initial assumption OiSP Opt > . Therefore, this case does not happen.
with respect
Case 2) OiM P OiSP holds. This relation is equivalent to OiM P LSP
i
= OiSP . Regarding OiM P LSP
to relation LSP
i , the stopping condition in master problem
i
M P (i) is satisfied and the master problems stops when it finds a feasible solution with an
upper bound UiM P satisfying the following relation.
SP
UiM P LSP
i = Oi
(EC.6.1)
(EC.6.2)
Similarly we can show that for k b(OiSP Opt)/c relation (EC.6.3) holds. This is because
it is supposed form iteration i to iteration i + b(OiSP Opt)/c all visited adversary scenarios
have repeated vectors w.
SP
SP
Oi+h
Oi+h1
h {1, 2, ..., k}
(EC.6.3)
h {1, 2, ..., k}
(EC.6.4)
(EC.6.5)
ec10
SP
O(i+h1)
Opt
OiSP Opt
(EC.6.6)
which is equivalent to
SP
Opt
O(i+k)
(EC.6.7)
Therefore, we proved that if in k = b(OiSP Opt)/c iterations after iteration i the algorithm
SP
does not find any adversary scenario with a repeated vector w then Oi+k
Opt holds.
ec11
(EC.7.1)
If the algorithm visits an adversary scenario with a repeated vector w in the subproblem
SP
regarding Lemma 2. Then with respect to
in iteration i + 1, we must have UiM P = Oi+1
SP
< Opt holds which is a contradiction because the optimal objective value of
(EC.7.1), Oi+1
the subproblem is an upper bound of the optimal objective of the robust problem. Therefore,
in this case in the next iteration an adversary scenario with a new vector w will be generated.
Case 2) OiSP Opt , OiSP OiM P and OiM P < Opt hold. We show in the next
iteration the algorithm generates an adversary scenario with a new vector w. Because of
OiSP OiM P , in the master problem in iteration i there is not any adversary scenario
satisfying the stopping condition. Thus, the master problem is solved optimally and we will
have the following relation.
OiM P = UiM P
(EC.7.2)
In the subproblem of next iteration, if the algorithm visits an adversary scenario with a
repeated vector w, then regarding Lemma 2 we must have relation (EC.7.3).
SP
UiM P = Oi+1
(EC.7.3)
Considering the primary assumption OiM P < Opt and relations (EC.7.2)- (EC.7.3) we must
SP
have Oi+1
< Opt which is a contradiction because the optimal objective value of the subprob-
lem is an upper bound of the optimal objective value of the robust problem. Therefore, in
this case in iteration i + 1 the algorithm generates an adversary scenario with a new vector
w.
Case 3) OiSP Opt , OiSP OiM P and OiM P = Opt hold. We show that in this case
in the next iteration either the Benders algorithm converges or it generates an adversary
scenario with a new vector w. Because of OiSP OiM P , in the master problem in iteration
i there is not any adversary scenario satisfying the stopping condition. Therefore, the master
problem is solved optimally and relation (EC.7.4) holds.
P
LM
= OiM P = UiM P
i
(EC.7.4)
ec12
solution of the robust problem is obtained and the Benders algorithm is converged. Therefore,
in this case, in the next iteration either the Benders algorithm converges or it generates an
adversary scenario with a repeated vector w.
ec13
(EC.8.1)
Note that g(i2 )1 g(i1 ) holds because i1 < i2 . Also regarding Lemma 2, in the subproblem
of iteration g(i2 ) that the algorithm has visited an adversary scenario with a repeated vector
SP
MP
. This relation together with (EC.8.1) demonstrates the validity
= Og(i
w, we have Ug(i
2)
2 )1
SP
SP
of Og(i
Og(i
.
2)
1)
ec14
n0
P
j=1
(1 + (b(OfSP
(j)+1 Opt)/c +
(EC.9.1)
For j = n0 we can use a similar reasoning as presented above for j < n0 . The difference is that
only Case 2 is applicable because regarding the definition of n0 no new vector w is visited
after visiting the n0 -th new vector w. Moreover, when we use Lemma 3 and 4 in Case 2,
the generation of an adversary scenario with a new vector w is not an option and we are
sure that after finding the n0 -th new vector w, the Benders algorithm converges in at most
b(OfSP
Opt)/c
+
2
iterations that is the same as (EC.9.1) with respect to If (j)+1 = 1
(j)+1
for j = n0 . Therefore, by summing the number of iterations computed by (EC.9.1) from j = 1
to j = n0 we obtain the following maximum number of iterations.
j
k
n0
n0
P
P
SP
0
1 + (Of (j)+1 Opt)/ + 1 If (j)+1 = n +
b(OfSP
Opt)/c
+
1
I
f
(j)+1
(j)+1
j=1
j=1
n0 +
n0
P
j=1
SP
0
SP
b(Of (j)+1 Opt)/c + 1 = n b(Og(1) Opt)/c + 2
SP
|W | b(Og(1)
Opt)/c + 2
ec15
Proof of the first inequality: We know that in n0 iterations the algorithm visits at least one
adversary scenario with a repeated vector w. g(1) denotes the iteration in which a repeated
vector w is visited for the first time. To prove the first inequality it is enough to show the
validity of the following relation (EC.9.2).
$ SP
%
$ SP
%
!
Og(1) Opt
Of (j)+1 Opt
+1
+ 1 If (j)+1
j {1, 2, ..., n0 }
(EC.9.2)
SP
As Og(1)
Opt is a valid relation, (EC.9.2) holds when If (j)+1 equal 0. In the case that
If (j)+1 is equal to 1, regarding the definition of g(1) and If (j)+1 we know that g(1) f (j) + 1.
SP
SP
SP
O(f
Thus, with respect to Lemma 5, we have Og(1)
(j)+1 that results in b(Og(1) Opt)/c + 1
b(OfSP
(j)+1 Opt)/c + 1. Therefore, relation (EC.9.2) is valid.
ec16
Figure EC.1
The adversary solution space in the example presented to show the non-optimality of the dual algorithm.
ec17
y,w
(EC.11.1)
dD
xd s
d D, s Sd
(EC.11.2)
dD
(EC.11.3)
dD
(EC.11.4)
dD
(EC.11.5)
tT
(EC.11.6)
wds {0, 1}
d D, s Sd
(EC.11.7)
ytp {0, 1}
t T, p Pt
(EC.11.8)
xd 0, integer
P
wds = 1
sSd
P P
ytp =
tT pPtd
swds
sSd
ytp = 1
pPt
It is clear that we can remove adversarys variables ytp and wds from the above model.
Also we can consider constraint (EC.11.2) only for the highest value s Sd for each d D.
Therefore, assuming that smax,d denotes the highest value s Sd , the non-adjustable nurse
planning problem reduces to the following model.
P
min
c1 xd
x
(EC.11.9)
dD
xd smax,d
dD
(EC.11.10)
xd 0, integer
dD
(EC.11.11)
The above model is trivial and its optimal solution is presented as follows.
smax,d
xd = max 0,
dD
(EC.11.12)
ec18
IF
OR
Sur.
Ave
Min
Max
STD
Ave
Min
Max
STD
1.1
1
2
3
4
5
39
79
119
157
202
2.63
5.00
7.46
9.26
11.69
2.14
4.29
7.14
8.21
9.57
2.93
5.86
8.07
10.86
13.36
0.22
0.46
0.30
0.80
0.89
3.15
6.20
9.10
12.07
15.19
2.79
5.50
8.64
11.07
13.43
3.71
7.36
9.71
13.57
16.21
0.28
0.50
0.36
0.79
0.71
1.3
1
2
3
4
5
39
79
119
157
202
2.63
5.00
7.46
9.51
11.89
2.14
4.29
7.14
8.00
9.86
2.93
5.86
8.07
10.36
12.71
0.22
0.46
0.30
0.72
0.78
3.15
6.20
9.10
12.07
15.19
2.79
5.50
8.64
11.07
13.43
3.71
7.36
9.71
13.57
16.21
0.28
0.50
0.36
0.79
0.71
1.5
1
2
3
4
5
39
79
119
157
202
2.65
5.02
7.51
9.58
12.01
2.14
4.50
7.14
8.07
10.00
3.00
5.86
8.50
10.50
13.29
0.24
0.43
0.45
0.72
0.80
3.15
6.20
9.10
12.07
15.19
2.79
5.50
8.64
11.07
13.43
3.71
7.36
9.71
13.57
16.21
0.28
0.50
0.36
0.79
0.71
1.7
1
2
3
4
5
39
79
119
157
202
2.67
5.06
7.54
9.76
12.49
2.14
4.50
7.14
8.36
10.14
3.00
5.86
8.50
10.86
13.79
0.24
0.43
0.43
0.70
0.88
3.15
6.20
9.10
12.07
15.19
2.79
5.50
8.64
11.07
13.43
3.71
7.36
9.71
13.57
16.21
0.28
0.50
0.36
0.79
0.71
1.9
1
2
3
4
5
39
79
119
157
202
2.67
5.06
7.54
9.75
12.89
2.14
4.50
7.14
8.43
10.36
3.00
5.86
8.50
10.71
15.43
0.24
0.43
0.43
0.66
1.24
3.15
6.20
9.10
12.07
15.19
2.79
5.50
8.64
11.07
13.43
3.71
7.36
9.71
13.57
16.21
0.28
0.50
0.36
0.79
0.71
Average
119
7.39
6.38
8.31
0.54
9.14
8.29
10.11
0.53
ec19
Table EC.12.2- Details of the number of first-stage nurses in the best and non-adjustable solutions
for instances with a planning horizon of three weeks (L = 3).
Data Info.
IF
OR
Sur.
Ave
Min
Max
STD
Ave
Min
Max
STD
1.1
1
2
3
4
5
59
121
182
240
300
3.36
6.09
8.71
11.31
14.17
3.00
5.05
7.76
10.67
13.43
3.76
7.52
9.52
12.81
15.10
0.20
0.67
0.53
0.56
0.54
4.31
8.64
12.61
16.57
20.71
3.90
7.48
11.57
15.67
19.43
4.71
9.86
13.19
17.33
21.95
0.23
0.63
0.48
0.56
0.70
1.3
1
2
3
4
5
59
121
182
240
300
3.36
6.28
8.95
11.73
14.66
3.00
5.43
7.95
11.14
13.90
3.76
7.19
9.76
13.00
15.38
0.20
0.53
0.52
0.51
0.52
4.31
8.64
12.61
16.57
20.71
3.90
7.48
11.57
15.67
19.43
4.71
9.86
13.19
17.33
21.95
0.23
0.63
0.48
0.56
0.70
1.5
1
2
3
4
5
59
121
182
240
300
3.48
6.67
9.25
11.78
14.90
3.14
5.76
8.48
11.14
14.05
3.90
7.95
9.90
13.14
15.52
0.23
0.66
0.42
0.55
0.52
4.31
8.64
12.61
16.57
20.71
3.90
7.48
11.57
15.67
19.43
4.71
9.86
13.19
17.33
21.95
0.23
0.63
0.48
0.56
0.70
1.7
1
2
3
4
5
59
121
182
240
300
3.50
6.86
9.90
12.30
15.90
3.14
6.29
9.14
11.52
14.24
3.9
7.81
10.57
14.00
20.86
0.24
0.46
0.44
0.72
1.77
4.31
8.64
12.61
16.57
20.71
3.90
7.48
11.57
15.67
19.43
4.71
9.86
13.19
17.33
21.95
0.23
0.63
0.48
0.56
0.7
1.9
1
2
3
4
5
59
121
182
240
300
3.50
6.88
10.83
15.70
20.40
3.14
6.24
9.14
12.14
19.19
3.90
7.81
12.67
17.00
21.52
0.24
0.47
1.16
1.43
0.66
4.31
8.64
12.61
16.57
20.71
3.90
7.48
11.57
15.67
19.43
4.71
9.86
13.19
17.33
21.95
0.23
0.63
0.48
0.56
0.70
Average
180
9.62
8.72
10.73
0.59
12.57
11.61
13.41
0.52
ec20
Table EC.12.3- Details of the number of first-stage nurses in the best and non-adjustable solutions
for instances with a planning horizon of three weeks (L = 4).
Data Info.
IF
OR
Sur.
Ave
Min
Max
STD
Ave
Min
Max
STD
1.1
1
2
3
4
5
80
163
241
318
397
3.67
6.79
9.45
12.13
15.33
3.21
5.93
8.36
11.43
14.68
4.18
7.54
9.89
12.93
16.00
0.31
0.44
0.45
0.46
0.43
5.01
9.88
14.34
18.74
23.09
4.36
8.86
13.71
18.18
22.18
5.39
10.71
14.71
19.64
23.79
0.26
0.56
0.31
0.46
0.48
1.3
1
2
3
4
5
80
163
241
318
397
3.83
7.13
9.77
12.74
15.82
3.46
6.21
8.57
12.00
15.14
4.18
7.96
10.32
13.79
16.46
0.22
0.48
0.50
0.52
0.41
5.01
9.88
14.34
18.74
23.09
4.36
8.86
13.71
18.18
22.18
5.39
10.71
14.71
19.64
23.79
0.26
0.56
0.31
0.46
0.48
1.5
1
2
3
4
5
80
163
241
318
397
3.96
7.47
10.10
13.33
19.44
3.57
6.50
8.79
12.29
15.61
4.29
7.86
10.79
14.57
23.14
0.23
0.38
0.61
0.70
3.16
5.01
9.88
14.34
18.74
23.09
4.36
8.86
13.71
18.18
22.18
5.39
10.71
14.71
19.64
23.79
0.26
0.56
0.31
0.46
0.48
1.7
1
2
3
4
5
80
163
241
318
397
4.00
7.78
10.75
16.10
22.84
3.57
7.21
9.54
12.64
22.00
4.29
8.64
11.54
19.39
23.57
0.19
0.43
0.53
2.78
0.46
5.01
9.88
14.34
18.74
23.09
4.36
8.86
13.71
18.18
22.18
5.39
10.71
14.71
19.64
23.79
0.26
0.56
0.31
0.46
0.48
1.9
1
2
3
4
5
80
163
241
318
397
4.01
8.52
14.04
18.44
22.79
3.57
7.32
13.39
17.71
21.96
4.29
10.50
14.46
19.39
23.54
0.18
1.25
0.35
0.50
0.47
5.01
9.88
14.34
18.74
23.09
4.36
8.86
13.71
18.18
22.18
5.39
10.71
14.71
19.64
23.79
0.26
0.56
0.31
0.46
0.48
Average
240
11.21
10.19
12.14
0.66
14.21
13.46
14.85
0.42
ec21
Table EC.12.4- Details of the number of first-stage nurses in the best and non-adjustable solutions
for instances with a planning horizon of three weeks (L = 5).
Data Info.
IF
OR
Sur.
Ave
Min
Max
STD
Ave
Min
Max
STD
1.1
1
2
3
4
5
101
202
302
401
503
3.76
6.71
10.04
13.04
17.06
3.51
6.46
9.46
12.49
16.49
3.94
7.11
11.00
13.89
17.86
0.17
0.21
0.44
0.41
0.40
5.29
10.33
15.31
20.19
25.30
5.11
10.00
14.74
19.54
24.49
5.49
11.03
16.06
20.89
26.31
0.11
0.28
0.41
0.44
0.55
1.3
1
2
3
4
5
101
202
302
401
503
4.03
7.29
10.44
13.7
17.51
3.71
6.86
9.80
13.34
16.89
4.26
7.66
11.26
14.57
18.14
0.16
0.26
0.45
0.37
0.40
5.29
10.33
15.31
20.19
25.30
5.11
10.00
14.74
19.54
24.49
5.49
11.03
16.06
20.89
26.31
0.11
0.28
0.41
0.44
0.55
1.5
1
2
3
4
5
101
202
302
401
503
4.30
7.51
11.59
17.98
25.03
3.89
7.03
9.91
13.74
24.37
4.66
8.23
15.69
20.4
26.17
0.22
0.37
2.05
2.68
0.55
5.29
10.33
15.31
20.19
25.30
5.11
10.00
14.74
19.54
24.49
5.49
11.03
16.06
20.89
26.31
0.11
0.28
0.41
0.44
0.55
1.7
1
2
3
4
5
101
202
302
401
503
4.23
7.93
11.88
19.99
25.02
3.97
7.57
10.8
19.34
24.31
4.54
8.77
15.63
20.66
26.00
0.16
0.36
1.36
0.45
0.52
5.29
10.33
15.31
20.19
25.30
5.11
10.00
14.74
19.54
24.49
5.49
11.03
16.06
20.89
26.31
0.11
0.28
0.41
0.44
0.55
1.9
1
2
3
4
5
101
202
302
401
503
4.26
8.89
15.03
19.96
25.01
4.06
7.57
14.40
19.34
24.34
4.46
9.97
15.63
20.66
26.00
0.13
0.97
0.39
0.48
0.52
5.29
10.33
15.31
20.19
25.30
5.11
10.00
14.74
19.54
24.49
5.49
11.03
16.06
20.89
26.31
0.11
0.28
0.41
0.44
0.55
Average
302
12.49
11.75
13.49
0.58
15.28
14.78
15.95
0.36