pasi 2013 santiago, chile integrated consolidation facility location and inventory routing for...
TRANSCRIPT
PASI 2013
Santiago, Chile
Integrated Consolidation Facility Location and Inventory Routing for Supply Networks
Ronald G. Askin [email protected]
with thanks to Mingjun Xia
School of Computing, Informatics, and Decision Systems Engineering
Arizona State University
PASI 2013
Santiago, Chile
Overview
• Introduction
• Problem Definition and Background
• Multi-Product Integrated Supply Chain Network Model
– Multi-product FLP with Approximated IRC Function
– Inventory Routing Problem
– Integrated Problem’s Results and Analysis
• Consolidation Facility Location & Demand Allocation
• Global Sourcing Options for Multistage Production
• Conclusion and Future Work
2
PASI 2013
Santiago, Chile3
Who’s That Speaker?
Ronald G. Askin, DirectorSchool of Computing, Informatics, and Decision Systems EngineeringArizona State UniversityTempe, AZ 85287-8809 [email protected]
PASI 2013
Santiago, Chile4
Arizona State University
PASI 2013
Santiago, Chile
The State of Arizona
PASI 2013
Santiago, Chile6
Background and Activities
• BS in IE, Lehigh University (in the time of punch cards)
• MS OR and PhD in Industrial and Systems Engineering, Georgia Tech
• Professor of Industrial Engineering
• Fellow of Institute of Industrial Engineers (IIE)
• Former IIE Board of Trustees Member
• Former Chair of Council of Industrial Engineering Academic Dept Heads
• Former President of INFORMS M&SOM Society
• Editor-in-Chief IIE Transactions
PASI 2013
Santiago, Chile7
IIE Transactions
The flagship journal of the Institute of Industrial Engineers and hopefully your preferred choice for publication.
http://www.tandfonline.com/toc/uiie20/current
• Methodological focus in most papers• Real world applications/impact• Original, innovative contribution required• Novel problems and models encouraged
PASI 2013
Santiago, Chile
Logistics: Facility Layout to Supply Networks
• Production Control Systems• Supply Chain Design• Batch Sizing/Lot Streaming
• Queuing Networks
• Material Flow & Capacity Models
• Facility Layout
PASI 2013
Santiago, Chile
Introduction
• Supply Chain Management (SCM)
9
The goal: To deliver the right product to the right place at the right time for the right price, while minimizing system-wide costs and satisfying service requirements.
PASI 2013
Santiago, Chile
Supply Chain Management Decisions
10
Strategic Level
Tactical Level
Operational Level
Long Term
Medium Term
Near Term
• Corporate objectives• Capacity / Facilities• Markets to operate• Location• Resources
• Aggregate planning• Resource allocation• Capacity allocation• Distribution• Inventory management
• Shop floor scheduling• Delivery scheduling• Truck routing
Facility Location Problem (FLP)
Inventory Control Problem (ICP)
Vehicle Routing Problem (VRP)
PASI 2013
Santiago, Chile
Motivation
• Lots of research has been done in each area in SCM, but few models comprehensively address the integrated network.
• To achieve a global optimal (or near optimal) solution, it is necessary to consider the entire system in an integrated fashion and include all trade-offs in a realistic fashion.
• We will look at the Distribution side (post production).
11
Customer service goals
Inventory Strategy• Forecasting• Inventory decisions• Purchasing and supply scheduling decisions• Storage fundamentals• Storage decisions
Transportation Strategy• Transport fundamentals• Transport decisions
Location Strategy• Location decisions• The network planning process
PASI 2013
Santiago, Chile
Our Distribution Problem
12
Made Here in Volume
PASI 2013
Santiago, Chile
Distribution Problem Scenario
13
Product Mixes Sold here by the Item at many Outlets
PASI 2013
Santiago, Chile14
Global Reality – But let’s start regionally
PASI 2013
Santiago, Chile15
Our Distribution Problem
•Assume (global) manufacturing system is defined.
•Goal: Distribute completed products to retail outlets.
•Assume goal is a (distribution) system optimal solution.
•Assume a relatively stable environment.
•Assume system to be designed from scratch – (any existing facilities could be sold for value or are on short-term leases).
PASI 2013
Santiago, Chile16
Planning Decisions
Where to place Distribution Centers?
How large to make DCs?
How to ship from Factory to DC – Quantity, frequency, form, mode?
How to take advantage of load consolidation opportunities?
How to serve each retail outlet – from where and how often?
How much safety inventory to keep and where to keep it?
PASI 2013
Santiago, Chile
What’s Relevant?
•Locations of Producers
What Else?
•Locations of Retailers•Cost of Transportation•Cost of Facilities by site/capacity(Fixed, Variable Operating)•Vehicle Capacities•Demands and Patterns•Product Substitutability•Inventory and Shortage Costs/Policies•Product Lifetime•Supply Dependability and Lead Times
PASI 2013
Santiago, Chile18
Important (Real-World) Factors Ignored
Insert your list here:
• Stochasticity of demand
• Dynamic nature of demand (multiple periods)
• Substitutability of products
• Strategic corporate initiatives (profit, service, competitiveness)
• Financial risk and return on investment
• Taxes, duties, exchange rates if multinational
• Reverse logistics (collection, refurbishment)
• Direct shipments
PASI 2013
Santiago, Chile
Facility Location Problem (FLP)
• Fermat-Weber (1909): A simple facility location problem in which a single facility is to be placed, with the only optimization criterion being the minimization of the sum of distances from a given set of point sites.
• More complex problems: the placement of multiple facilities, constraints on the locations of facilities, and more complex optimization criteria.
• The goal: to pick a subset of potential facilities to open, to minimize the sum of distances from each demand point to its nearest facility, plus the sum of fixed opening costs of the facilities.
• The facility location problem on general graphs is NP-hard to solve optimally, by reduction from (for example) the Set Cover problem.
• Daskin (2002), Ozsen (2008): include inventory control decisions in FLP.
19
PASI 2013
Santiago, Chile
Inventory Control Problem (ICP)
• Harris (1913): Economic order quantity (EOQ)• Clark and Scarf (1960): Multi-echelon Inventory
• Inventory control: the supervision of supply, storage and accessibility of items in order to ensure an adequate supply without excessive oversupply.– Where to hold inventory?– When to order?– How much to order each time?
• The goal: the order quantity and the reorder point are determined such that the total cost is minimized.– Total cost = Purchasing cost + Setup Cost + Holding Cost + Shortage Costs
• The single-item stochastic inventory control problem is NP-hard even in the case of linear procurement and holding costs. (Halman et al. , 2009)
20
PASI 2013
Santiago, Chile
Vehicle Routing Problem (VRP)
• Dantzig and Ramser (1959):To deliver goods located at a central depot to customers who have placed orders for such goods.
• The goal: to minimize the cost of distributing the goods.
• The vehicle routing problem in general is NP-hard as it lies at the intersection of these two NP-hard problems: – Traveling Salesman Problem – Bin Packing Problem
• Inventory Routing Problem (IRP): An extension to include inventory concerns.Kleywegt, Nori, Salvesbergh, Transportation Science, 2002
21
PASI 2013
Santiago, Chile22
How Good are the Models?
“A conclusion that can be drawn from the literature devoted to the uncapacitated facility location problem and its extensions is that the research field has somehow evolved without really taking the SCM context into account. Features … have been included in the models in a rather general way and specific aspects, that are crucial to SCM, were disregarded. In fact, extensions seem to have been mostly guided by solution methods.”
- Melo et al. EJOR, 2009
PASI 2013
Santiago, Chile23
Principles to Keep in Mind
1. Pooling Synergy Safety Stock
Di
Di
Di
Di
Di
Di
1/2( )iZ Var D 1/2( )iZ Var D 1/2( )iZ Var D
1/2( )iZ N Var D
Assumes independence
PASI 2013
Santiago, Chile24
Principles to Keep in Mind
2. Inventory vs. Service Level
What’s the Traditional Perspective?
Inventory
Fill Rate
100%
PASI 2013
Santiago, Chile
Comment on Second Principle: Little’s Law
In Steady State,
Average Inventory = Consumption Rate x Ave. Time in System
N = XT or L = λW
Diminishing Returns: Beyond the “elbow” more inventory is just more cost and more opportunity for degradation, loss, congestion and cost!
PASI 2013
Santiago, Chile26
But it’s even worse
Beyond a threshold increasing inventory reduces sales!
• Congestion slows service response
• Inventory is outdated
• Forecast horizons too long for accuracy
Carburetors vs. Fuel Injection
PASI 2013
Santiago, Chile
Empirical Profile: Know When Enough is Enough
0
2
4
6
8
10
12
0 10 20 30 40 50 60 70 80 90
Th
rou
gh
put
WIP
Little's Law and Chaos
Deterministic
Exponential
Empirical
Remember
L =λW
N = XT
In theory, there’s no difference between theory and practice, in practice there is. – Yogi Berra
PASI 2013
Santiago, Chile
Multi-Product Integrated Supply Chain Network Model
28
PASI 2013
Santiago, Chile
Existing Research is Helpful but Not the Same
• Shen, Z.J.M, Qi, L., 2007. Incorporating inventory and routing costs in strategic location models. European Journal of Operational Research 179, 372-389.– Single Producing Plant (One Supplier, One Product)– Uniformly located customers across an area– (Q,r) inventory ordering/replenishment model for DC– Fixed and identical routing frequency from DC to customers– Single routing tour from each DC (1 vehicle) for cost estimation– Uncapacitated DCs
• Javid, A.A., Azard, N., 2010. Incorporating location, routing and inventory decisions in supply chain network design. Transportation Research Part E 46, 582-597. – Single product, no transhipment– (Q,r) inventory model for DC– Known delivery route frequency– Fixed vehicle capacity per year
29
PASI 2013
Santiago, Chile30
Our Problem (Model)
Integrated supply chain network design: location, transportation, routing and inventory decisions
Multi-product and plant supply chain network.
Transshipments between DCs.
Non-uniformed (clustered) customer locations.
Multiple routes with model-determined frequencies from DCs.
Nonlinear inventory costs (safety stock).
Full truck load deliveries to DCs with choice of truck size.
PASI 2013
Santiago, Chile
Consolidation and Transportation Centers
31
DistributionCenter
Retailer
Facility
B: Only Distribution Center
ConsolidationCenter
A: Through Consolidation Center
Retailer
Facility
Facility
C: Point-to-Point Transportation
Retailer
DistributionCenter
Fixed Location Cost V.S.Transportation savingsEasy management
PASI 2013
Santiago, Chile
Multi-Product Integrated Supply Chain Network Model
32
10
11
6
7
8
9
4
5
1
2
31
2
3
4
5
6
Facility DC Retailer
1
2
3
4
Distribution RetailerConsolidationFacility
1
2
3
4
5
6
1
2
3
4
5
6
7
8
9
10
11
PASI 2013
Santiago, Chile
Problem Description
– Location decision: How many DCs to locate, where to locate, how much capacity at each opened site.
– Transportation decision: Allocate facilities and retailers to opened DCs.
– Routing decision: Routing tours and frequencies to retailers.
– Inventory decision: How often to reorder, what level of inventory stock to maintain.
33
10
11
6
7
8
9
4
5
1
2
31
2
3
4
5
6
Facility DC Retailer
1
2
3
4
PASI 2013
Santiago, Chile
Problem Description
Facility – DC – Retailer:
– Each production facility supplies a single product.
– Retailers are clustered in the service region.– Demand follows a known stationary
distribution. – Single source: all products at one retailer
should be delivered by one DC.– Full truck load (FTL) shipping is used from
plants to DCs and between DCs, multiple truck size choices exist.
– Routing delivery is used for shipment from DCs to retailers.
34
10
11
6
7
8
9
4
5
1
2
31
2
3
4
5
6
Facility DC Retailer
1
2
3
4
PASI 2013
Santiago, Chile
Research Scope and Activities
35
Two-phase: Phase I: Multi-product FLP with Approximated IRC Function DC Locations and plant/retailer assignments
Approximate cost function for routing delivery cost (Shen, Z.J.M, Qi, L., 2007).
Phase II: Inventory Routing Problem Routing tours and frequencies
Solve the routing problem for each open DC and retailers assigned to it.
1
2
3
4
10
11
6
7
8
9
4
5
1
2
31
2
3
4
5
6
Facility DC
Retailer
1
2
3
4
Retailer
DCFacility
1
2
3
4
5
6
1
2
3
4
5
6
7
8
9
10
11
1
2
3
4
RetailerDCFacility
1
2
3
4
5
6
1
2
3
4
5
6
7
8
9
10
11
PASI 2013
Santiago, Chile
Cost Components
FC: Annualized fixed cost of locating DCs
SC: FTL Shipping cost from plants to DCs and between DCs
36
10
11
6
7
8
9
4
5
1
2
31
2
3
4
5
6
Facility DC Retailer
1
2
3
4
Qpj
Qpj’j
' '
'' ,
' '
' ' ' '
'
'' , ''
pj j pi pj jii I
pj pi pjj ij J i I
pj l pjl j j l j jll L l L
pj pjl pjl l pjl j j j jl j jl l j jll L l L
pj jpj p Pj pj j jp P j J j j
pj j j
Q Y
Q Y
q q T q q T
A a b q T A a b q T
QQSC A A
q q
, jk jkj J k KFC f O
p j, j’ i
Q = total shipped/timeq = quantity per trip (shipping mode capacity)A = cost per tripT = if use that truck size
PASI 2013
Santiago, Chile
Cost Components
SSC: Safety stock inventory holding cost at DCs
RIC: Regular inventory holding cost at DCs
37
10
11
6
7
8
9
4
5
1
2
31
2
3
4
5
6
Facility DC Retailer
1
2
3
4
Qpj
Qpj’j
2 2' ' ', ' ' , 'pj pj pi pjj i j j pi pj jii I j J j J j j i I
SS z lt Y lt Y
' '
' , ''2 2
pj j j pj jpj j J j j
p p pj jp P
q q QRI
v v Q
Ypjj’I binary route indicator variables
PASI 2013
Santiago, Chile
Phase I: Multi-product FLP with Approximated IRC Function
38
PASI 2013
Santiago, Chile
Approximated IRC Function
IRC: Annual inventory routing cost from DCs to retailersrji a computationally estimated parameter to represent the annual
IRC at retailer i if assigned to DC j
39
, ji jii I j JIRC r y
max | ( ) |ji jn R i
( , ) ( ). . :
jji lml m A i
s t d D
( ), j
pm pm R i p P
ji
vq
Inventory costRouting
cost
Nearest insertion method
5
4
DC
1
23
6
7
8
9
10
11
12
1min
2ji pi ji
ji n pi pip Pn Nji n n
a cf h z
n f f s
PASI 2013
Santiago, Chile
Approximated IRC Function
IRC: Annual inventory routing cost from DCs to retailersrji a computationally estimated parameter to represent the annual
IRC at retailer i if assigned to DC j
, ji jii I j JIRC r y
direct shipping method
21min 2
2pi ji
ji ji n pi pip Pn Nji ji
da cd f h z
s
/ 2ji ji jir
PASI 2013
Santiago, Chile
Problem Formulation
1 jij JY i I
Minimize (FC + IRC + SC + SSC + RIC)Subject to:
Single source Single path
Link variables
Maximum number of PWsSingle capacity level
Truck size selection
Throughput Capacity limit
'',1 ,pj jij j J
Y i I p P
', '
,pj ji jip P j JY MY i I j J
', ' ,pjj i pji I j J
Y MW p P j J
pj pj J
W PW p P
1 jkk K
O j J
1 ,pjll l
T p P j J
' 1 , ' , 'jj ll L
T j j J j j
', , , ' , '
pi ji pi pjj i jk jkp P i I p P i I j J j j k KY Y C O j J
' ', , , , , {0,1} , , ' , ,jk ji pj ji pj pjl jj lO Y Y W T T i I j j J p P k K
PASI 2013
Santiago, Chile
Single Plant Warehouse Case (n = 1)
Optimal truck size from plants to PWs and transshipment between DCs
DSRIC: direct shipping and regular working inventory holding cost
TSRIC: Transshipment and regular working inventory holding cost
42
min2
pii I lpj pjl pjl l pjp Pl
l
qDSRIC a b q h
q
' '' ' '
'
min2
pj jp P pj jlj j j jl j jl l pjp Pl
l pj jp P
Q QqTSRIC a b q h
q Q
Shared transhipment loads
PASI 2013
Santiago, Chile
Nonlinear Terms
Safety stock
Working inventory
43
2 2' ' ', ',pj pi pjj i j j pi pj jii I j J i I
lt Y lt Y
2 2 2 2' ' ', '
' , ''
2', '
2' ' '
pj pi pjj i j j pi pj jii I j J i Ipj j J j j
pj pj j
pj pj pi pjj ii I j J
pj j j j pi pj jii I
z lt Y z lt YSS
S S
S z lt Y
S z lt Y
'
'
' '
pj j
pj jp P
pj j pi pj jii I
Q
Q
Q Y
Recursive procedure to update safety stock parameters, Gebennini et al. (2009)
Same holding rate
PASI 2013
Santiago, Chile
Tabu search can avoid search cycling by systematically preventing moves that generate the solutions previously visited in the solution space. Simulated annealing allows the search to proceed to a neighboring state even if the move causes the value of the objective function to become worse, and this allows it to prevent falling in local optimum traps.
Construction stage• Greedy method• Minimizing initial Fixed Cost (FC)• Minimizing initial Inventory Routing Cost (IRC)
Improvement stage• Location improvement
• Close an opened DC; Open a closed DC• Assignment improvement
• Assign one retailer to another reachable DC• Assign one PW to another opened DC
44
Tabu Search-Simulated Annealing (TS-SA)
PASI 2013
Santiago, Chile
Fixed Cost (FC) and Inventory Routing Cost (IRC) are two major cost components
FC 24% Vs. IRC 50% (Shen and Qi, 2007)
FC Heuristic: Minimizing initial FC• Set covering problem: minimizing total number of DCs• To include cost consideration
Open all necessary DCs Open additional DCs to save total cost
Improvement stage: TS-SA
45
Ad-Hoc Heuristics
PASI 2013
Santiago, Chile
Inventory Routing Cost (IRC) is a major cost componentIRC accounts for 50% of total (Shen and Qi, 2007)
IRC Heuristic: Minimizing initial IRC Open all DCs and assign retailers to its nearest DC Close unnecessary DCs to save total cost
Improvement stage: TS-SA
46
Ad-Hoc Heuristics
Nothing new, sounds like variable selection in regression.
PASI 2013
Santiago, Chile47
Lower Bound: Without considering transshipment between DCs.
,
'
'' , ''
' '2 2' ' ', , ' ' , ' ' , '
'
= +
2 2
j pj pj pjj J p P j J
pj jpj p Ppj j jp P j J j j
pj j j
pj j j pj jpj pj pi pjj i j j pi pj jip P j J i I j J j J j j i I j J j j
pj jp P
pj
SC h SS RI
QQA A
q q
q q Qh z lt Y lt Y
Q
A
2', , '
2
, , '
,
2
, '
+2
=2
Where min
pj pjpj pj pi pjj ip P p P j J i I j J
pj
pi pj pji Ipj pj pj pi pjp P j J i I j J
pj
pj pjp P j J
pi pi Ipj l L pjl pjl l pj pj pii I j J
l
Q qh z lt Y
q
h qA h z lt W
q
W
ha b q h z lt
q
2j lq
n =1
PASI 2013
Santiago, Chile
• All points (plants, DCs, and customers) are geographically dispersed in a 500 * 500 miles region.
• Plants are randomly distributed, retailers are clustered into m groups with the centers of gravity also randomly distributed in this space.
• 8 different data sets with each set including 15 scenarios with sizes ranging from 20 to 200 retailers and 5 to 20 products.
• Data sets differ in fixed location cost rate (low, high), demand rate (case 1, case 2) and holding cost rate (low, high).
• All the computational times are obtained on a Intel(R) Core(TM)2 T5550 at 1.83 GHz using Windows 7. Three introduced heuristics are applied in Microsoft Visio Studio C++. IBM ILOG CPLEX Optimization Studio is used to solve the modified model and lower bound model.
48
Parameter Settings
Retailers Case 1 Case 2
High demand 10% consume 27% TD 10% consume 80% TD
Medium demand 80% consume 70% TD 10% consume 10% TD
Low demand 10% consume 3% TD 80% consume 10% TD
PASI 2013
Santiago, Chile49
Partial Results
PASI 2013
Santiago, Chile50
Partial Results
PASI 2013
Santiago, Chile51
Results Comparison
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 105 110 115 1204.5
4.7
4.9
5.1
5.3
5.5
5.7
5.9
6.1
6.3
Greedy TSSA IRC-TSSA FC-TSSA LowerBound
Log of Cost (to show separation)
PASI 2013
Santiago, Chile52
Large Problems 200 Retailers, 20 DCs
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100 105 110 115 120200000
400000
600000
800000
1000000
1200000
1400000
1600000
Greedy TSSA IRC-TSSA FC-TSSA LowerBound
PASI 2013
Santiago, Chile
• Computational time: Maximum scenario: 1863 seconds by heuristics V.S. No solution after 5 hours of computation for even some small instances.
• Objective values: the greedy solution’s objective value is reduced by 29.1% on average ( the improvement is greater under large instances). 12.1% higher than the lower bound (do not include transshipment consideration and large instances do not converge completely in CPLEX).
• IRC and FC heuristics: better than simple TSSA method, especially in large instances, indicating the importance of a good starting point.
53
Results and Analysis
PASI 2013
Santiago, Chile
• IRC heuristic performs the best in both the number of best solution scenarios and the average GAP.
54
Results and Analysis
TSSA IRC FC
Best Solution Scenarios
34 of 120 66 of 120 47 of 120
Average GAP 11.6% 2.0% 3.1%
Cost %
PASI 2013
Santiago, Chile55
Results and Analysis
FC: fixed location cost; IRC: inventory routing cost;IC: regular inventory and safety stock cost at DCs; SC: total shipping cost
FCRate = Low; HCRate
= Low
FCRate = Low; HCRate
= High
FCRate = High;
HCRate = Low
FCRate = High;
HCRate = High
0%10%20%30%40%50%60%70%80%90%
100%
21.9% 16.2% 35.9% 27.9%
57.4% 63.4% 47.1% 54.3%
6.9% 7.4% 5.4% 6.9%
13.7% 12.9% 11.7% 10.9%
SCICIRCFC
PASI 2013
Santiago, Chile
Phase II: Inventory Routing Problem
56
PASI 2013
Santiago, Chile
Phase II: Inventory Routing Problem
Inventory Routing Problem (IRP)
• Inventory management and transportation are two of the key logistical drivers of supply chain management.
• Bottom problem of the integrated supply chain network design.
57
PASI 2013
Santiago, Chile
Literature Review
• The classification criteria used for the IRP
•Missing/Rare in IRP research – Uncertain demands at retailers– Variable routing frequency– Optimal number of vehicles– Nonlinear characteristics of routing cost and lead time
……
58
Characteristic AlternativesTime Instant Finite InfiniteDemand Stochastic DeterministicTopology One-to-one One-to-many Many-to-many
Routing Direct Multiple ContinuousInventory Fixed Stock-out Lost sale Back-order
Fleet composition Homogeneous Heterogeneous
Fleet size Single Multiple Unconstrained
Products Single Multiple
PASI 2013
Santiago, Chile
Problem Description
• Inventory routing problem– One distribution center (DC) and multiple retailers (R).– Each retailer has an independent demand for the product, follows normal
distribution.– The DC uses homogeneous capacitated vehicles for routing delivery.– Routing frequencies fall in a discrete set such as daily, every other day,
etc.
• Decisions– Routing tour to each retailer. – Routing frequency of each tour.
• Total cost– Routing cost over each trip: predetermined fixed cost + a variable cost
depending on total distance.– Inventory cost at each retailer: cycle inventory + safety stock.
59
PASI 2013
Santiago, Chile60
Problem Formulation
(6)
number of trips for route v in one yearv n vnn Nf Z
0,v ij ijvi j Rd d X
distance of route v
1 v viv Vr
v viv V
d Rlt
R s
lead time for retailer r. Lead time is a function of routing route frequency (first component) and route distance (second component)
Then the objective function is to Minimize:
Routing cost and inventory cost. Inventory at each retailer includes both cycle inventory to meet foreseeable demand and safety stock to overcome uncertain demand.
0.5 pip P
v v pi pi rv V i R p P v Vvi vv V
a cd h z ltR
PASI 2013
Santiago, Chile61
Problem Formulation
Vehicle capacity
Every route starts from DC
Flow conservation
Distance capacity
Single assignment
Subtour elimination
0 0
0
0
,
0
0 ,
1
| | | | 1 , ,
,
1
,
viv V
pi vip P i R
v
v
sv tv stv
stv tsvs R s R
iv stvi R s t R
itv vit R
vnn
R i R
Rq v V
d D v V
M M R X R s t R v V
X X t R v V
R X X v V
X R i R v V
Z
0
1
, , {0,1} , , , ,
0 ,
N
vi stv vn
iv
v V
R X Z i R s t R v V n N
M i I v V
Variable connection
Route frequency
PASI 2013
Santiago, Chile
• Evidence indicates that the sweep method for routing vehicles is computationally efficient and produces an average gap from optimality of about 10 percent.
• Modify the simple sweep method by considering specific characteristics in this problem– Optimize routing tour after inserting a new retailer point.– Optimize routing frequency within one route.– Start from each retailer and sweep both clockwise and counterclockwise.
62
Modified Sweep (MS)
5
4
DC
1
23
6
7
8
9
10
11
12
PASI 2013
Santiago, Chile
• Distance between two routes: The smallest possible distance between two retailers in two distinct routes. Let Sk be the set of retailers included in route k; Dij be the distance between retailer i and j, and DRmn be the distance between route m and n, then:
• Adjacent route: Two routes are called adjacent if the distance between these two routes is smallest compared to other routes (or within some predetermined value).
Move 1: Exchange their delivery order within one route.Move 2: Exchange two retailers from two adjacent routes.Move 3: Insert one retailer to an adjacent route.Move 4: Open a new individual route for one retailer.
63
Tabu Search-Simulated Annealing (TS-SA)
arg min{ } , mn ij m nDR D i S j S
PASI 2013
Santiago, Chile
• MethodologyGenerate an initial solution where each retailer is serviced by one individual tour, and then try to merge retailers into one route.
• Natural frequencyThe optimal routing frequency for each retailer under an individual tour.
ILS1: Fixed vehicle cost + variable cost from DC to the retailer. ILS2: Fixed cost + Variable cost (twice the distance from the DC)ILS3: Fixed cost + Variable cost (Distance limitation)] / Average number of retailers in one route
64
Integrated Local Search Method (ILS)
Whether to merge two retailers depends on two factors: the distance between these two retailers and similarity in natural frequency.
PASI 2013
Santiago, Chile
Procedures:
1. Calculate natural frequency for each retailer.
2. Divide all retailers into different groups. In this research, four groups will be generated with routing frequency to be 350, 175, 50, and 25, respectively. G1, G2, G3 and G4.
65
Integrated Local Search Method (ILS)
5
4
DC
1
23
6
7
8
9
10
11
12
PASI 2013
Santiago, Chile
3. Use embedded modified sweep method to merge retailers in group G1.
4. Try to insert other retailers in other groups (in the order of G2, G3 and G4) in current routes.
5. Repeat the same process of step 3 and 4 for retailers in group G2, G3 and G4 respectively.
(6). Improvement step: Tabu search. This step is not necessary.
66
Integrated Local Search Method (ILS)
DC
32
1
13
Route 1
Route 2
DC
PASI 2013
Santiago, Chile67
Integrated Local Search V.S. Modified Sweep
BA
60 mile
100 mile
60 mile
200 mile
r4
r3r2
r1
B
200 mile
100 mile
100 mile
150 mile150 mile
200 mile
200 mile
200 mile200 mile
150 mile
100 mile
A
200 miler4
r3
r2
r1
ILS method provides a better solution
MS method provides a better solution
PASI 2013
Santiago, Chile
• Fixed partition policy (FPP)The retailers are partitioned into disjoint and collectively exhaustive sets. Each set of retailers is served independently of the others and at its optimal replenishment rate.
• Use a genetic algorithm (GA) to generate/update a fixed partition for all retailers.
• A TSP is solved within each partition and optimal delivery frequency is selected accordingly.
68
Hybrid Genetic Algorithm Method (HGA)
Generate fixed partitions
Find optimal delivery tour (TSP)
For each fixed partition
Decide optimal delivery frequency
Calculate total cost of each fixed partition
Update the best fixed partition
Is the termination criterion satisfied
STOP
YES
NO
PASI 2013
Santiago, Chile
• Objective Values The major benefit of routing comes from reduction in delivery cost. In an ideal case:– Delivery distance to one retailer is 1 + 1/(N+1) times the distance between nearest neighbors.– Smallest total number of trips required is total demand over all retailers dividing by a truck
capacity.
Alternatively, consider delivery distance to each retailer as D/n, where D is the distance limit and n is the average number of retailers in one route.
Any feasible solution is an upper bound, a simple solution is using all direct-shipping.
69
Bounds
,
11
1 1 11
1 2
rpr prr R p P p P
LB r r r prr R p Pr r
dN
IRC a c d h zC N p
1
2
prp Pe r r prr R p P
r r
a cD DIRC h z
n np
PASI 2013
Santiago, Chile70
Parameter Settings
Name Notation Value Remark
Service level zα 1.96 97.50%
Vehicle capacity C 150Distance limit D 500 milesVehicle speed s 500 miles/dayFixed cost a $ 5/truckVariable routing cost cd c = $ 0.1mile d = distance (miles) Available frequency/year
fn {25, 50, 175, 350} 1year = 350 days
Location of DC 0 (0, 0)Number of retailers N {20, 50, 100, 150, 200}Locations of retailers (x, y) [-100, 100] Uniform Distribution
Demand mean/year 10% Low: [50, 150] 80% Medium: [500, 2000]10% High: [10000, 25000]
Uniform Distribution
Demand deviation/year Low: [1, 5] High: [10, 50] Uniform Distribution
Holding cost hr Low: $ 10/unit year Medium: $ 50/unit yearHigh: $ 100/unit year
r
r
PASI 2013
Santiago, Chile71
Results and AnalysisCPU Seconds Cost in $1,000
PASI 2013
Santiago, Chile
•All heuristics except HGA work well in terms of objective values. •Using routing strategy can reduce total cost by 25.8% - 51.4%.• When the holding cost and demand variance decrease, the benefits of routing
strategy also decreases. • Routing strategy will have more benefits if the demand or optimal order size of
each retailer is small compared to vehicle capacity.
72
Results and Analysis
hr σr N = 20 N = 50 N = 100 N = 150 N = 200 Average
High High 37.4 44.2 45.5 47.9 48.4 44.7High Low 42.1 44.6 48.1 51.4 51.3 47.5Medium High 29.7 40.5 42.4 42.1 48.6 40.7Medium Low 29.9 32.3 38.6 38.5 42.6 36.4Low High 29.6 33.7 37.5 37.2 37.6 35.1Low Low 25.8 32.2 35.9 37.5 37.2 33.7
Savings vs. Direct (Individual) Delivery
PASI 2013
Santiago, Chile
• Among all heuristics, Modified Sweep method performs the best and HGA method is the worst. – Modified sweep method already captures many important aspect of this routing
problem. – With capacity and distance constraints, there is a high probability that a child
from crossover and mutation is infeasible, especially in large instances.
• ILS works very fast in terms of CPU time, but its objective values are much higher than MS.
• If joint with Tabu search, ILS-TS generates better results than MS in large scenarios, but CPU time increases because of Tabu search step.
73
Results and Analysis
Use MS method for IRP in this research stage, and we can also use TS to further improve results from MS method if necessary.
PASI 2013
Santiago, Chile
Integrated Problem’s Results and Analysis
74
PASI 2013
Santiago, Chile
Integrated Problem’s Results and Analysis
Two-phase solution approach: • Phase I: Multi-product FLP with
Approximated IRC Function• DC Locations and retailer/plant
warehouse assignments– TS-SA method
• Phase II: Inventory Routing Problem • Routing tours and frequencies: solve the
routing problem for each open DC and retailers assigned to it.– MS method ( + TS)
75
Original network design problem
Phase I: FLP with approximate IRC
DC locations and retailer assignments
Phase II: IRP, route assignments
Each opened DC
PASI 2013
Santiago, Chile
• Heuristics work well in terms of objective values compared to the original greedy solution. The original greedy solution’s value is reduced by 25.3% on average.
• The IRC gap in the table is calculated as |Real IRC/Approximated IRC - 1|. The average value for this gap is 5.6%.
• Reasons for using
76
Results and Analysis
/ 2ji ji jir
Using average cost 5.6%
Using only routing cost 14.4%
Using direct shipping cost
32.8%
(1 )
,
ji ji jir w w
D Cw f
d
Demand density
Locationdensity
PASI 2013
Santiago, Chile
Consolidation Facility Location and Demand Allocation Model
77
PASI 2013
Santiago, Chile
Consolidation Facility Location and Demand Allocation Model
• Each production facility ships its product directly to each opened DC.
• Two shipment methods: direct shipment from facility and indirect shipment through a DC.
• Product sets: consolidation is allowed for shipping products in the same product set.
• Each facility provides one specific product
• Single route for each product
• Retailers hold safety stock only if the lead-time of replenishing one order is above a specific threshold value
78
PASI 2013
Santiago, Chile79
Problem formulation
2 42
4 2
3 1 5
5 3 3 1 1
11 1 1 1
1 1
2
2
ik iki
kj kj ik ir ik ikrk j k k i r
s skr r ir ir
k s iir ir kr ikr kr ir ir ir ir
r r i k
ir ir irir ir ir ir
i ir
Q hf w z h lt x
Q h Q hz h t x lt t x lt
Q D xA a bl Q
C Q
22 2 2 2
2 2
33 3 3 3
3 3
ir ikrik r
ik ik ik ikr k iir ik ik
s ir ikrskr i S
kr kr kr kr sk r s kr kr
D xQ
A a bl QC Q
D xQ
A a bl QC Q
Fixed cost
Transportation cost
Inventory costs at retailers
Inventory costs at DCs
PASI 2013
Santiago, Chile80
Problem formulation
Shipment mode
DC’s capacity level
Total shipping quantity of one product set
Open DCs before assignment
2
3
3 3
222
1 2 3
,
, ,
, ,
1
1 ,
0.8 2
, , 0 , , ,
, 0,1
ik kjj
skr kj
j
skr ikr
i S
kjj
ikr irk
ikir ik ikr kj kj
i i r j
sir ik kr
ir ikr
Q Mw i k
Q Mw s k r
Q Q s k r
w k
x x i r
Qlt x U w k
Q Q Q i s k r
x x
, ,i k r
DC’s capacity constraint
PASI 2013
Santiago, Chile81
Genetic Algorithms
i k r
irx
ikrx
i1 k1 r1
i2 k2 r2
Product 1 (P1) Product 2 (P2)From i to
rFrom i to
k From k to r From i to r From i to k From k to r
1 0 1 0 0 1 0 0 0 0 1 0 1 1 0 00 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Shipment directions
Feasible solution example
Chromosome representation
PASI 2013
Santiago, Chile
• Build the solution step by step using a “cascade” method.
Procedures:• Set t = 0.• To build a table with “I R” rows and “K+1” columns and evaluate the feasibility. If
feasible, then calculate the objective function for each product-mode-retailer combination OFikr(t) comparing the K+1 possibilities of shipment (directly by plant or by K DCs). Otherwise put the OFikr(t) equal to a big integer called M.
• Comparing the value of OFikr(t) for each row to select the minimum and the second smallest for each row respectively called Minir= mink{OFikr(t)}, SecMinir = mink {OFikr(t)/Minir}.
• Calculate Δir as the difference between Minir and SecMinir (potential regret).
• Select the maxir{ Δir } and in correspondence to the column, fix the solution for the relative product/retailer couple. Set t = t +1.
• Repeat the steps 2-5 for “I R” iterations.
82
Greedy Construction Heuristic
PASI 2013
Santiago, Chile
• Two different product sets in the tests. • Production facilities and retailers are chosen as major cities in the United
States. Potential DCs can be located at the locations of retailers. • Each DC has three possible sizes: small, medium and large. • Fixed location cost: home value, Daskin (1995).• Capacity: potential service amount.• Demands of products: normally distributed. The mean is proportional to the
population around that retailer. The variance of demand is calculated using coefficient of variation times mean demand.
• Lead time: distance.• Shipping cost: fixed cost + variable costs (distance, shipping quantity).
83
Parameter Settings
PASI 2013
Santiago, Chile84
Scenarios construction
Scenario No.Plants
No.DCs
No. Retailers
Function to define the batch size
Length of Chromosome
1 2 10 10 Floor 1402 2 10 10 Ceiling 1403 5 10 10 Floor 6004 5 10 10 Ceiling 6005 2 10 49 Floor 6086 2 10 49 Ceiling 6087 5 10 49 Floor 27458 5 10 49 Ceiling 2745
PASI 2013
Santiago, Chile85
Results and Analysis
GeneticHeuristic
(Max {Delta = SecMin – Min})Heuristic (Min {Min})
All Direct-
shipping
Time (sec)
No.Iter.
Obj. Value
No. DCs
Time (sec)
No.Iter.
Obj. Value
No. DCs
Time (sec)
No.Iter.
Obj. Value
No. DCs
Obj. Value
1 98 422 3.057E+8 4 1 20 3.030E+8 3 1 20 3.030E+8 3 4.211E+8
2 123 329 3.055E+8 4 1 20 3.030E+8 3 1 20 3.030E+8 3 4.211E+8
3 1435 9238 1.181E+9 4 3 50 9.827E+8 6 3 50 9.846E+8 7 1.180E+9
4 1203 1399 1.180E+9 4 3 50 9.791E+8 7 2 50 9.846E+8 7 1.180E+9
5 80 652 5.060E+7 0 10 98 5.060E+7 0 10 98 5.060E+7 0 5.060E+7
6 39 589 5.062E+7 0 10 98 5.062E+7 0 9 98 5.062E+7 0 5.062E+7
7 17771 23760 1.743E+8 1 49 245 1.545E+8 0 25 245 1.545E+8 1 1.545E+8
8 20486 25900 1.515E+8 1 49 245 1.545E+8 1 24 245 1.545E+8 1 1.545E+8
PASI 2013
Santiago, Chile
• The heuristic proved computationally efficient and provided the best solution in all but one case (scenario 8).
• The “delta” form of the heuristic (making the selection based on difference between the best and second best options) outperformed the “min” form in two cases and the “min” form performed best in one case.
• Both heuristics perform at least as well as direct shipments.
• GA found the unique best feasible solution in the last case but requires significantly longer computation time.
86
Results and Analysis
PASI 2013
Santiago, Chile
• Proposed an innovative framework for a multi-product integrated supply chain network design problem.
• Derived and evaluated the effectiveness of a two-phase solution methodology.
• Transshipment is allowed between DCs, and routing delivery strategy is considered.
• Heuristics are generated in each phase to find a good solution in a reasonable time.
• Phase I: TS-SA method with an initial solution starting minimizing total IRC.• Phase II: MS method.
• Only the special case of the original problem where only one PW is allowed for each plant is discussed in detail in current research.
87
Conclusion
PASI 2013
Santiago, Chile88
Future Needs
• Better understanding of need for integrated formulations
• Better models for iterating/decomposing between levels (?)
• More realistic models – (problem driven not analysis driven)
• Integration of production and distribution decisions
• Taxonomy of actual problems by industry, logistics method
• Expansion to stochastic optimization (two-stage)
PASI 2013
Santiago, Chile
Thanks!
Q & A
89