Partha Datta

Multi-agent model for a complex
supply chain: Case of a Paper
Tissue Manufacturer
by
Partha Datta
Martin Christopher &
Peter Allen
Cranfield University School of Management
Contents
•Complex Systems & Supply Networks
• Need for new supply chain modelling framework
• Agent Based Modelling Framework
• Case Study
• Application of the Framework – Results
• Conclusion
• Contribution
Complex systems & Supply Networks
Complex Systems
• Consist of different interacting elements,
• The elements may be very different and change with
time
• The elements have some degree of internal
autonomy
Supply Networks
• A supply chain is a network of organizations
• Firms in seemingly unrelated industries can compete
for common resources
• Firms keep on moving in and out of network
• Firms have own decision making ability
Complex systems & Supply Networks
Complex Systems
• Elements are coupled in a non-linear fashion
• Behavioural patterns created through myriads of interactions
Supply Networks
• A small fluctuation at the downstream can cause large
oscillations upstream (BULL-WHIP)
• Collective behaviours emerge beyond the control of any single
firm
Existing supply chain modelling
techniques
• Existing network planning tools are deterministic
• Optimization models are offline and brittle
• Strongly focus on physical transactions
• Investigate various supply chain activities in an isolated
way
• Historically modelling has been top-down
• Abstraction and assumptions limit representing reality
- None of these approaches is rich enough to
capture the dynamical behaviour of the entire
supply network
Need for a new modelling framework
• Is bottom-up, starts by identifying the most basic
building blocks – the agents
• Should be able to model the independent control
structures of each agent
• Should be able to model the mutual attuning of
activities based on interdependence
• Should reveal and aim to integrate the material
structure, the information structure, the decision
structure and the strategic structure
Agent Based Modelling [ABM]
• Provides a method for integrating the entire supply
chain as a network system of independent echelons
(Gjerdrum et al, 2001)
• Can represent many actors, their intentions, internal
decision rules and their interactions (Holland, 1995
and 1998; Axelrod, 1997; Prietula, 2001)
– Agents have some autonomy
– Agents are interdependent
– Agents follow simple rules
Agent Based Model Building Blocks
Network KPIs
Network Average Inventory per SKU, CSL,
Production set-up costs
Decision Making Stage
Determine RDC preferences for dispatching
materials, Determine SKU preference for
production - both based on a combination of
forward cover and inventory turnover (to
avoid over-forecasting errors), Inventory
targets based on CSL
Internal KPIs
Average Inventory, CSL, Sales
backlogs
Variables & Parameters
Sales,
Forecast,
Production Capacity,
SKU list,
Lead Time,
Packing constraints
Customer
Orders
Functioning Stage
Customer
Agent
Order
Queu
Order
Management
Delivery
Queue
Delivery
Management
Production Planning
& Control
FGI
Inventory Planning
CSL: customer service level, FGI: Finished Goods Inventory
Figure 1: The General Agent structure used in the model
Production
Goods
Inward
Supplier
RDC Agent
or Factory A
Agent Based Model Building Blocks
Production Factory agent
• Decision Making Stage –
– 1.Target finished goods inventory determination
– 2.Ranking of products for determining priority for
production
• Functioning Stage –
–
1. Production, Planning & Control : based on
the forecast demand during approximate production
time window, fixed production rate for each product,
–
2. Palletisation & Delivery : delivery to central
warehouse in specified pallet types
Agent Based Model Building Blocks
Distribution centre agents
• Decision Making Stage –
– 1.Safety and Target Stock Determination,
– 2.Replenishment Policy Adoption,
• Functioning Stage –
Order Management : aggregates all demands,
forecasts
– 2. Goods Dispatch Management : availability
based partial fulfilment of orders
– 3. Finished Goods Inventory Management :
replenishment of inventory based on target
inventory and reorder point levels based on safety
stock levels estimated at decision making level
–
1.
Case Study – A Paper Tissue
Manufacturing Company
Delay Objects
load +
ship
Distribution Centre
Agents
FLINT
UK order
bank
S2
Repal to
S2
load +
ship
NIEDERBIPP
load+ship
LOGIS
E3
CH order
bank
CZ order
bank
E3
load+ship
Factory Agent
Koblenz Factory
E3
E5
RUSSIA
RU order
bank
E3
Central
Warehouse at
Koblenz
order bank
DE/NL/BE/CH/Nordic
+ DDXM France
E3/E5
E5
load+ship
Distribution
Centre Agent E5
load+ship
VSE
FR order
bank
Marene
IT order
bank
E5
load+ship
Arceniega
ES/PT order
bank
E5
load+ship
Fig. 9. Supply Chain structure under study
[ E3/E5/S2 are the pallet sizes ordered by the countries]
EDE
order bank
CH
Customer
Agents
The Complex Supply Network - Details
• Varying lead times for different countries
• Different pallet size requirements
• Different product portfolio requirements
• Some products are demanded by single country
• Different products have different demand patterns
• All products share the same machine resource for
production
• Different products have different times of set-up
Bottlenecks –
• “Marketing driven” production – not “market
driven”
• Mismatch between real demand and forecast
- Higher repalletisation costs
- Lack of balance in production
- Correct products not in stock at right place
• No common KPIs
Data
• Forecast and Sales data collected during period from
1st January to 31st December 2004
• Forecast data is monthly and Sales is approximated
by the daily delivery amounts
• Data on daily inter-company deliveries and delivery
to customers are collected
• Theoretical and Empirical distributions are fitted to
the sales data to generate replications for simulation
Additional Data
•
•
•
•
•
Production Rates
Production Categories for change-over
Change-over times
Swiss Sales Data
Maximum and Minimum Production Cycle Times for
some products
• Pallet Size Constraints
• Product, Market, Supplier, Pallet-size combination
• Delivery Lead Times
Applying the framework
The functioning and decision making stages
• Rationing and priority based on increasing order size
• order backlogs have the highest priority
• Ordering is based on forecast, forecast error, stock
position and forecast bias
• Order quantity is decided based on each RDC agent’s
- knowledge of central warehouse stock
- perception of stock wear out and demand variability
• Use of global information for allocating time for
production
• Priority for production is decided based on
- forward cover of product codes in RDCs and central
warehouse
- absorptive power of product codes
Model Validation
• The difference between Modelled (83838) and Actual
(84124) Total Average Network Inventory across 8
codes for the stipulated time period (for which actual
data was obtained) found to be within 0.34% of
Table 8b: Validation Results - Production Figures
Actual.
Product Code
Table 8a: Validation Results - Inventory Figures
Product Code
RDC
RDC Average
Inventory
Actual
Model
741
751
Difference
1.35%
Average Production Amounts
Actual
Model
Difference
Wypall7122
298
290
2.68%
Wypall7126
94
94
0.00%
Wypall7122
UK
Wypall7198
Koblenz
19784
19879
0.48%
Wypall7190
533
473
11.26%
Wypall7122
Niederbipp
195
175
10.26%
Wypall7196
44
48
9.09%
Kimcel7025
France
309
312
0.97%
Wypall7198
366
322
12.02%
Wypall7190
Italy
4032
3487
13.52%
Wypall7341
343
308
10.20%
Wypall7342
117
131
11.97%
Performance Measures
• Customer Service Level (CSL)
TH
 AS
CSL =
t 1
TH
D
t 1
n ,t
and
ASn,t = min (In,t-1,Dn,t)
n ,t
Where, ASn,t = actual sales in simulation n at time instance t
Dn,t = demand in simulation n at time instance t
In,t = ending stock level in simulation n at time t
n = simulation number
TH = simulation time horizon
• Production Change-Over
• Average Inventory at each regional distribution
centre
• Total Network Inventory
Model Performance Vs Actual System Performance
(Over-all/Global performance)
• The model shows improved inventory and CSL
performance in a balanced manner across the supply
chain
• The total number of changeovers is 80 as compared
to 132 in actual case
• The model idle time = 22 days, actual system idle
time = 47 days
• Repalletisation Modelled value = 197379 as
compared to actual value of 202606, a reduction of
2.6%
• The model also produced better balance in allocating
total production time across codes with respect to
actual demand
Conclusion
• Firm's operations must be driven by current
customer requests
• Methodology to understand the key issues essential
for improving operational resilience in a complex
production distribution system
- knowing earlier
- managing-by-wire
- designing a supply network as a complex system
- production and dispatching capabilities from the
customer request back
Contribution
• Studies and provides methods for improving the
management of uncertainty and thereby improving
resilience in complex multi-product, multi-country
real-life production distribution system
• Provides a generic agent-based computational
framework for effective management of complex
production distribution systems.
Scope for further research
• Use of market data to include effects of competition
in different country markets
• Extension to include raw material supply chain
• Inclusion of cost data to understand various tradeoffs
Why Supply Chain Management is so
difficult?
• Nonlinearities –
1. Reliance on forecasts at each stage for basing decisions
2. Different demand patterns of different products over time
3. Different constraints (lot-sizing, transport capacity etc.)
4. Different supply chain structures
• Results into upstream demand amplification (Bull-whip)
Actual demand, actual average stock and actual
total time of production at Koblenz
Actual Stock Levels
Actual Stock Levels at Koblenz and Ede for
product X9
The information and material flow - Actual
Sales Forecast
Actual
Customer
Order
Nomenclature
Central Planning
Product Specifications
Monthly RDC
stock plan
Yearly
production
budget
Processes
Order
acceptance
Storage points
Tasks/Actions
Order Bank
Stages in process
interdependencies
Rough
planning
basesheet
production
Fine planning
basesheet
production
Rough
planning
converting
Inventory
control
BASESHEET
PRODUCTION
Rough
Planning
Transport
Fine
planning
converting
CONVERTING
Mill
Basesheet
Stock
Koblenz
Basesheet
Stock
Distribution
Fine Planning
Transport
particular actions
Current
Stock
Distribution
Rough
Planning
Transport
Fine Planning
Transport
Changing Premises of Industrial Organisation
Source: www.dti.gov.uk
Modelled System vs Actual System Performance
Actual Sales and Modelled Stock Levels for product X12 at
UK RDC
stock in number of
cases
2000
1500
1000
500
0
1
22
43
64
85 106 127 148 169 190 211 232 253 274 295 316 337 358
tim e in days
Modelled Stock
Actual demand
Modelled System vs Actual System Performance
Stock at Koblenz
Balance in Factory
A Complex System includes the “system you
see” and the hidden processes that change it
This is not just asking how a system runs, but WHY it exists. It must express
synergetic behaviour of its components in that environment:
Beginning
System 1
1 type
Structural
Change
occurs...
Instabilities
System 2
2 types
System 3
4 types
System 4
8 types
System 5
6 types
Later...
A “Complex System” creates and destroys
transitory traditional Systems…..
Time
Production Planning & Control
Decision making stage of
the agent
Target finished goods
inventory and ranking for
production priority
[Finished goods
inventory/ total forecast]
for each product ranked
> 1
time for
producing the
top-ranked
product
t
h
tp
th > tp
yes
no
th=tp
produce for the
calculated time
period
change-over time
for a certain timeperiod (CO)
CO>CO*
yes
no
continue
production
produce products
according to stipulated
maximum and minimum
time periods
Flowchart 2: Production, Planning & Control
forward cover of all
products in next
stockpoint arranged
in ascending order
If top ranked
product is
produced within
the past 7 days?
Top ranked
product's forward
cover < 0
Yes
Yes
No
No
time, each
product is last
produced
Do not consider the
product for production for
7 days
If top ranked
product is
produced within
the past 4 days?
No
target finished goods
inventory of all
products in next
stockpoint
If top ranked product
target finished goods
inventory in next stockpoint >0
No
Yes
Yes
Start producing the topranked product
Yes
cumulative sales of
all products in the
network
total stock of all
products in the
next available
stock-point
If top ranked product
cumulative sales until a
particular time-period
>0
No
Start producing next ranked product for
which the above are non-zero and
positive and the cumulative sales/total
inventory at next available stock-point is
the highest
Do not consider the
product for production for
4 days