1
A Cut Approach for Solving the Single- Row Linear Layout Problem
SIAM 2014: Session CP-30 July 07~11, 2014
Chicago, IL, USA
Shine-Der Lee
Department of Industrial & Information Management
National Cheng Kung University
1 University Road, Tainan, 70101, Taiwan
sdlee@mail.ncku.edu.tw
1. Problem Description
Scenario:
A family of parts is processed in the system
The routing sequence of parts are known and fixed
Input/output stations are located at two ends of the single- row linear layout
Forward flow: part moves from left (i.e., Input station is at the left side) to right (Output station is at the right side)
Backward flow: part moves from right (Output station) to left (Input station), also called backtracking
Per unit backtracking movement cost can be different from that of forward flow
Decision model:
Find the layout sequence or design (a permutation of
workstation sequence) so that the total material flow cost for a given routing requirement is minimized.
3
1. Problem Description (continued)
WS n
Linear flow path
WS 1 WS i
Input station 0
Output station n+1 WS j
S0
( ) j 1 α = + β
Material flow cost is the sum of part flow multiplied by the distance between any two workstations
Equal-distance between any two adjacent locations
Bi-directional material handling system, e.g., AGV, conveyor, guided rail, etc.
n workstations are to be assigned
backtracking
2. Some Established Results
Classical layout models: Quadratic Assignment Problem (QAP)
Koopmans and Beckman (1957): QAP
Lawler (1963): Integer programming model
Sahni and Gonzalez (1976): NP-complete
Optimal Linear Ordering problem (OLO)
Adolphson and Hu (1973): Network flow property
Garey and Johnson (1976): NP-complete
Single-row linear layout
Kouvelis et al. (1996): Linear integer programming model, minimize the times of backtracking, branch & bound
Palubeckis (2012): Improved lower bound, optimization approach with Tabu search
5
2. Some Established Results
Single-row linear layout (Alternative objective functions)
Picard and Queyranne (1981): Dynamic programming
Houshyar and McGinnis (1990): Minimizing WIP, cut tree based heuristic
Sarker et al. (1995): Multi-Pass Heuristic and Depth-first Insertion Heuristic
Samarghandi et al. (2010): Continuous version of the linear layout, Particle optimization heuristic
Open issues:
Minimizing the total material flow cost
Exact procedures with improved computational efficiency
Solving realistic-size problems, e.g., over 100 work stations for assembling iphones, and nearly 300 work stations for ipads.
3. The network flow Model
Objective function:
(1)
: The set of workstations that are located before the (q+1)
stsequence in the linear layout
: The set of workstations that are located after the q
thsequence in the linear layout
: The cut value (total material flow cost) for the corresponding q-locale network, e.g., see Picard
& Ratliff (1978), it can be expressed as:
(2)
0 , 1
1 1 1 1
( ) ( ) ( ) ( 1) 1 ( ) ( )
2
n n n n
i i n ij
i i i j
F α t α i t + α i n w α i α j
= = = =
∑ ∑ ∑ ∑
= + − + + −
{
| ( )}
Sq = j α j ≤ q
{
| ( )}
Sq = j α j > q ( , )
q q q
C S S
( , )
, 1q q q q
q q q oi i n ij
i S i S
i S j S
C S S t t
+w
∈ ∈
∈∑ ∑ ∑ ∑∈
= + +
3.1 The q-locale network
7
An example: The equivalent transformation for q = 1 locale network
q-locale network: The sub-network where the minimum cut is used to determine the qth assignment (location vector) of the workstation
Output station (n+1) Input
Station(0)
1
2
n i
, 1
tn n+
t01
t02
t0,i
t0, n
1,n 1
t +
2,n 1
t +
, 1
ti n+
w12
w2, n
,
wi n
w1, n
w1,i
w2, n
( )
1 1 1 0 1, 1 1
2 2
, = n i n n i
i i
C S S t t + w
= =
∑ + + ∑
3.2 Equivalence properties
Proposition 1: Let be one of the minimal cut in the q-locale network and be the corresponding
location vector, there exists a minimum cut and layout sequence such that:
for any workstation j, 1 ≤ j ≤ n, and
Proof: By the definition of cut value on the q-locale network and using term-by-term comparison.
(S Sq′ ′, q)
(S Sq, q) α′
α *
* ( ) ( ) 1, ,
* ( ) ( ) 1, ,
* ( ) ( ),
q q
q q
j j if j S S
j j if j S S
j j otherwise
α α
α α
α α
′ ′
= − ∈ ∩
′ ′
= + ∈ ∩
= ′
(
*) ( )
q(
q,
q)
q(
q,
q)
F α − F α ′ ≤ C S ′ ′ S − C S S
3.2 Equivalence properties
Proposition 2: There exists a set of q-locale networks
with the minimum cut , and , for q = 1,. . . , n-1, where α is the
optimum layout sequence with the minimum total flow cost, and vice versa.
Proof:
From (1), the objective value of a layout vector can be partitioned into three parts. From (2), the minimum cut for the q- locale network corresponds to the minimum flow cost of the layout vector for the workstation considered at this location.Hence, the set of minimum cuts of the q-locale network, for q = 1, . . , n-1, n, corresponds to the minimum cost for the optimum layout. Conversely, using Proposition 1, we can find a minimum cut with a reduced cut value, which implies that the minimum
cut is the solution corresponding to the optimal layout. 9 ( , )
q q q
C S S Sq =
{
j | ( )α j ≤ q} {
| ( )}
Sq = j α j > q
10
4. An Efficient Solution Procedure
The cut approach for linear row layout
1. Let S
q= {0}, i.e., it includes only the input station.
2. For q = 1, find the minimum cut for the 1-locale network as follows:
Generate a network (graph) using the input station as source node and the output station as sink node. All n
workstations are placed in between, i.e., every workstation is a candidate to be assigned to the first location. The arc capacity for any two nodes i and j is the part flow in
between.
Compute n cut values:
for n workstations. Find the minimum cut value.
Assign workstation i* in this minimum cut to the qth layout sequence. Tie breaking is arbitrary.
(
,)
, 1q q q q
q q q oi i n ij
i S i S
i S j S
C S S t t + w
∈ ∈
∈∑ ∑ ∑ ∑∈
= + +
11
4. An Efficient Solution Procedure
The cut approach for linear row layout (continued) 3. Update . Increase q by 1 and repeat steps
2 and 3 until q = n-1. Go to the next step.
4. The optimum layout sequence has been found;
and the total minimum flow cost is
Complexity analysis: To find the optimum layout, the total computations for the q-locale networks are: . The complexity of the cut approach is , where n is the problem size.
{ }*
q q
S = S i
1
0 , 1
1 1 1
( ) ( , )
n n n
i i n q q q
i i q
F
α
t t + − C S S= = =
=
∑ ∑
+ +∑
α
( )
1
2 1
n
i
i i
=
⋅ −
∑
21 1
( 1) (2 1) ( 1)
2 6 2
n n
i i
n n n n n
i i
= =
⋅ + ⋅ + ⋅ +
= ⋅∑ −∑ = −
( 3) Ο n
4.1 An illustration
Consider a bi-directional single-row linear layout with 6
workstations to be assigned. Flow between workstations is given below, where 0 is the input station and 7 is the output station.
0 1 2 3 4 5 6 7
0 -- 20 5 0 12 0 14 0
1 0 -- 24 12 0 15 0 0
2 0 0 -- 10 0 0 15 14
3 0 19 0 -- 0 12 0 10
4 0 12 10 14 -- 0 0 15
5 0 0 0 0 29 -- 10 12
6 0 0 0 5 10 24 -- 0
7 0 0 0 0 0 0 0 --
13
4.1 An illustration
The flow matrices for the q-locale network are as follows:
1. Flow matrix T from/to I/O
2. Flow matrix W
1 2 3 4 5 6 0 20 5 0 12 0 14 7 0 14 10 15 12 0
1 2 3 4 5 6
1 -- 24 31 12 15 0
2 24 -- 10 10 0 15
3 31 10 -- 14 12 5
4 12 10 14 -- 29 10
5 15 0 12 29 -- 34
6 0 15 5 10 34 --
4.1 An illustration
Computational details of the cut approach for the 1- locale network
0
1
q = S1={2, 3, 4, 5, 6}
1 {1, 3, 4, 5, 6}
S =
1 {1, 2, 4, 5, 6}
S =
1 {1, 2, 3, 5, 6}
S =
1 {1, 2, 3, 4, 6}
S =
*
1 {1, 2, 3, 4, 5}
S =
0 6 7
4.1 An illustration
Computational details for q = 4 and 5
15 Workstation 1 is assigned to the 4th sequence
Workstation 3 is assigned to the 5th sequence. Thus, the only unassigned workstation 2 is assigned to the last
sequence
4.1 An illustration
The optimum layout vector is α = (6, 5, 4, 1, 3, 2)
The minimum flow cost of the layout is the sum of:
Flow cost from the input station to every workstation, = 20+5+12+14 = 51 (From flow matrix T)
Flow cost from every workstation to the output station, = 14+10+15+12 = 51 (From flow matrix T), and
Flow cost from the minimum cuts, = 101 + 135 + 135 + 143 + 101= 615 (Between workstations)
= 717
.
Verification of total flow cost ,
= 160 + 154 + (149 + 95 + 70 + 79 + 10) = 717
6 0 1
i i
t
∑=
6 7 1
i i
t
∑
= 51
( , q)
q q
q
C S S
∑
=( )
6 0 0 6 7 7 6 61 1 1 ,
1
i i i i ij ij
i i i j i
j
F α t d t d w d
= = = ≠
=
∑ ∑ ∑ ∑
= + +
17
5. Concluding remarks
Some findings:
The single-row linear layout problem can be transformed into the equivalent network optimization problem of
finding minimum cuts for the corresponding q-locale networks.
The minimum cut in the q-locale network is the optimal assignment of workstation in the qth location sequence, and vice versa.
An efficient solution procedure, i.e., a cut approach, of complexity is developed to solve this layout
problem.
Extension:
The cut approach can be extended to solve the non- equal distance linear layout problem.
( )3
Ο n