View
228
Download
1
Embed Size (px)
Citation preview
Closed-loop Supply Chain decisions using Game Theoretic Particle Swarm Optimisation
Presented by:
Kalpit Patne
4th Year UG Student,Department of Industrial and Systems Engineering,Indian Institute of Technology (IIT) Kharagpur, INDIA
Presentation Outline Introduction Problem Definition Solution Methodology
◦ Background of Key Concepts◦ Improved Particle Swarm Optimisation
Numerical Analysis Results and Comparison Future prospects
Introduction
Source: Google images
Closed-loop supply chains are designed and managed to explicitly consider the reverse and the forward supply chain activities over the entire life cycle of the product.
Forward chain
Reverse Chain
Introduction
New Product
Used ProductProducer Consumer
Fun Fact!
Source: Google images
Introduction
Source: Google images
Key considerations in optimizing the Supply Chain Network
Location and distance Current and future demand Inventory level Size and frequency of shipment Warehousing costs Transportation costs Mode of transportation
Introduction
The significant aspects we aim to address in this study-
1. Retailer Facility Location
2. Customer Zone Allocation
3. Optimal Pricing Policy
4. Optimal Inventory Cycle Time
Problem Definition
1. Retailer Facility Location
Customer zones and production centre locations are known We want to optimally locate retailer facilities
Problem Definition
Source: Google images
2. Customer Zone Allocation
We want to know the customer zone allocation schema
Problem Definition
Source: Google images
3. Optimal Pricing Policy
We need to decide on the selling price of the new product and the rebate price (incentive) paid to the customer for returning the used product, so that the overall profit is maximized.
Since the demand and willingness to return is sensitive to selling and rebate price respectively, there exists a binding constraint on their values.
Problem Definition
4. Optimal Inventory Cycle Time
Deciding optimal inventory replenishment time
This has to be done considering the return shipments of used products from retailer facilities to the production centre for remanufacturing.
Problem Definition
Mathematical Representation of the Problem
Z= Max --------------------------------I
------------------------II
---------------------------------------------III
----IV
I – Profit function for new productsII – Profit in terms of assets retrieved from customersIII – Establishment and Operational cost of the facilityIV – Ordering and Inventory holding costs
Problem Definition
We have considered the proposed model as a combination of two sub-problems:
Location-Allocation Problem (LAP)1. Retailer facility location2. Customer zone allocation
Pricing-Inventory Problem (PIP)3. Optimal pricing policy4. Optimal inventory cycle time
Solution Methodology
Solution Methodology
LAP solved using modified
PSO
• No. of customer zones• Their locations• Production facility location• No. of retailer facilities
• Retailer locations
• Operational facilities
• Customer allocation
• Capacity• Fixed cost• Operational
cost• Distance
parameters
PIP solved with Gradient
search
• Optimal pricing policy
• Optimal inventory cycle time
1. Particle Swarm Optimization (PSO)
Memory based (personal and global bestposition)
Communication involved(for searching the global bestposition of the swarm)
Primary Algorithms used
Source: Google images
1. Particle Swarm Optimization (PSO)
Involves two main search variables- Position and Velocity
Updated in each iteration – fitness value for updated position of particle keeps on increasing with iteration
Drawback- If a particle discovers a local optimal, all the other particles will move closer to it, then particles are in the dilemma of local optimal point. ◦ This is known as premature convergence.
Primary Algorithms used
2. Evolutionary game-based replicator dynamics
It allows the fitness to incorporate the distribution of population types rather than setting the fitness of a particular type constant.
The general form of the evolutionary game consists of three key components: Players, Strategies space and Payoff function.
Particles in a swarm ------- Players in a gameResearch space ------- Strategies spaceVelocity ------- Rate of proportion changeFitness function ------- Payoff functiongBest ------- Evolutionary stable strategy
Proposed Solution Approach
Based on: Liu, Wei-Bing, and Xian-Jia Wang. "An evolutionary game based particle swarm optimization algorithm." Journal of Computational and Applied Mathematics 214.1 (2008): 30-35.
2. Evolutionary game based replicator dynamics
Where .xi – change rate of proportion of population of type i xi - proportion of population opting for strategy i, x=(x1,…,xn) - vector of the distribution of population types, fi(x) - fitness of type i (which is dependent on the population), ɸ(x) - average population fitness
Proposed Solution Approach
2. Evolutionary game based replicator dynamics
We update the velocities of particles using the replicator equation and then update the positions.
This speeds up the convergence of our algorithm.
Proposed Solution Approach
To avoid premature convergence, we use the concept of mutation.
Mutation- To diversify the population when it starts converging at a point.
A larger research space is explored and chances of obtaining a global optima increases
Diversity index, D(t)- Variable responsible to trigger mutation
Improved PSO (IPSO) Algorithm
Based on: Lin, M. O., and Zheng Hua. "Improved PSO algorithm with adaptive inertia weight and mutation." Computer Science and Information Engineering, 2009 WRI World Congress on. Vol. 4. IEEE, 2009.
Proposed Solution Approach
IPSO Adaptive Weight Update
In our algorithm we have used adaptive weight update method to make the updating process more dynamic.
We use the rate of change of fitness at each iteration to update the weight.
Proposed Solution Approach
Initialization
Iterations
If t > TWeight update
(linear)
Weight Update(adaptive)
Mutation trigger
Probabilistic mutation of some particle positions
FalseTrue
Yes
Velocity and Position Update (PSO)
Velocity and Position update (Replicator
Dynamics)
Update pBesti
Update gBest
No
We use a 100x100 2D space as our search space.
We used real life data where possible; otherwise, realistic assumptions were made
Numerical Analysis
Parameter Value# Customer Zones, M 20# Retailer Facilities, N 10
Numerical Analysis
This is a randomly generated market area showing the locations of production facility and all the identified customer zones
Results obtained after implementation of the IPSO algorithm
Numerical Analysis: Results
Performance of IPSO algorithm
Numerical Analysis : Results
Once the LAP problem is solved, we will have the retailer locations and customer zone allocation schema determined.
We then tackle the PIP problem and try to maximize the profit function Z (the mathematical model developed earlier) using gradient search method.
In our numerical example, we assumed:◦ manufacturing cost ‘c’ for a new unit = $100;◦ remanufacturing cost ‘cr’ = $15; and ◦ salvage value ‘s’ of the returned product = $60.
Numerical Analysis : Results
The gradient search algorithm produces the optimal value of selling price as p =127.57 for a new product and the incentive r=18.89 for returned products.
So, the maximum profit that the company can make by selling one unit of new product is (127.57 -100) = 27.57 per unit.
If the company decides to sell the remanufactured product at it’s salvage value, it will earn (60-18.89-15) = 26.11 per unit
Numerical Analysis : Results
The optimal inventory time Tj, is calculated from the standard Economic Order Quantity (EOQ) model equation.
Tj =
Numerical Analysis : Results
Retailer Facility (j)
Optimal Inventory Cycle Time (Tj)
1 Inf2 6.953 11.424 11.425 5.036 4.317 2.818 7.669 4.12
10 10.41
We have compared the results produced by our algorithm with three other algorithms: PSO, simulated annealing (SA), genetic algorithm (GA) to evaluate the
◦Performance (e.g. solution quality); and
◦Efficiency (e.g. convergence speed)
of the proposed algorithm.
Results and Comparison
Results and Comparison
0 5000 10000 15000 20000 25000 30000 35000200
400
600
800
1000
1200
1400
1600
IPSOPSOSAGA
Number of function evaluations
Fitne
ss v
alue
1. For M=20 and N=10
Results and Comparison
0 5000 10000 15000 20000 25000 30000 35000250
270
290
310
330
350
370
390
410
430
IPSOPSOSA
Number of function evaluations
Fitne
ss v
alue
1. For M=20 and N=10
(excluding GA)
Results and Comparison2. For M=30 and N=15
0 20000 40000 60000 80000 100000 120000400
600
800
1000
1200
1400
1600
1800
2000
2200
IPSOPSOSAGA
Number of function evaluations
Fitne
ss v
alue
Results and Comparison
0 5000 10000 15000 20000 25000 30000 35000450
470
490
510
530
550
570
590
IPSOPSOSA
2. For M=30 and N=15
Fitne
ss v
alue
Number of function evaluations
(excluding GA)
Results and Comparison3. For M=50 and N=20
0 5000 10000 15000 20000 25000 30000 350000
500
1000
1500
2000
2500
3000
3500
4000
IPSOPSOSAGAFit
ness
val
ue
Number of function evaluations
Results and Comparison3. For M=50 and N=20
0 5000 10000 15000 20000 25000 30000 35000760
780
800
820
840
860
880
900
920
940
IPSOPSOSA
Number of function evaluations
Fitne
ss v
alue
Further sensitivity analysis to improve robustness
Routing decision making can also be incorporated into the model
Salvage value of the returned products can be differentiated based on the age of used product
Further Improvements
SMART Infrastructure Facility Prof Pascal Perez Ms Tania Brown Dr Nagesh Shukla
School of MMM Prof Gursel Alici Prof Kiet Tieu Dr Senevi Kiridena
Prof Roger Lewis (ADR, EIS)
Acknowledgment
Q & A
Thank You!