optimal control strategy for relief supply behavior...
TRANSCRIPT
Optimal Control Strategy for Relief Supply Behavior after a Major Disaster
Riki Kawase
BinN International Research Seminar, Sep. 23, 2019
Doctoral Student, Department of Civil Engineering, Kobe University
Joint work with Dr. Urata and Prof. Iryo
Background: Humanitarian Logistics2
Humanitarian Logistics is important for minimizing the
damage between the rescue and restoration period after a disaster.
Difficulties
Material flow
Information flow
⋮⋮
⋮
Relief suppliers(RS)
Distribution centers(DCs)
Shelters
■ Node bottleneck on supply network (e.g., DCs don't work)
■ Link bottleneck on material network (e.g., Cannot access)
■ Link bottleneck on information network (e.g., Cannot communicate)
Supply DemandBuffer
3Chairs lined up to read "Paper, bread, water, SOS“
Transshipment processed inefficiently by human power
Many damaged roads
Kumamoto Earthquake (2016)
Background: Control Strategy4
Control Strategy in Japan
[Push-mode support (sequence control)] [Pull-mode support (feedback control)]
Demand forecasting Demand feedback
Pre-stock Push Pull
timePast disasters
Number of evacuees and meals supplied after the Kumamoto Earthquake
Meals
Evacuees
■ Long push-mode support caused the gap between supply (meals) and demand (evacuees).
Push Pull
Research Question
When should the control strategy change from push to pull ?
Background: Humanitarian Logistics5
Material flow
Information flow
⋮⋮
⋮
Relief suppliers(RS)
Distribution centers(DCs)
Shelters
Supply DemandBuffer
■ DCs can adjust gaps by holding inventories on the implicit assumption of supply and information availability.
Upper logistics Lower logistics
⁃ not available
⁃ limited available
⁃ fully available
⁃ Push-mode
⁃ Decentralized Pull-mode
⁃ Centralized Pull-mode
Push
Pull
border
Purpose and Methodology6
Purpose: Mathematical properties of the optimal control strategy
Research Question: When should the push- be changed to pull-mode ?
Methodology
■ Focus on information availability
■ Mathematically analyze the optimal push-mode (no information) and the optimal pull-mode (decentralized/centralized information).
■ Dynamic optimization approach using the stochastic optimal control theory considering Demand uncertainty.
■ The Bayesian updating process can model two control strategies.
updating interval =
∞ not available
0~∞ (limited available)
0 fully available
■ We drive the sufficient condition to change control strategy from push-mode to pull-mode.
7
Modeling
Definition8
𝐶𝑜𝑛𝑡𝑟𝑜𝑙 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒
𝑆𝑖𝑗(𝑡): Supply rate from node 𝑖 to node 𝑗 at time 𝑡
𝑆𝑡𝑎𝑡𝑒 𝑣𝑎𝑟𝑖𝑎𝑏𝑙𝑒
𝐼𝑖 𝑡 : Inventory level in node 𝑖 at time 𝑡𝐼𝑁 𝑡 : Net inventory level in the shelter at time 𝑡
𝑃𝑎𝑟𝑎𝑚𝑒𝑡𝑒𝑟
𝑟𝑖𝑗: Lead time from node 𝑖 node 𝑗(𝑟13 < 𝑟12 + 𝑟23 = 𝑟123 )
𝑧(𝑡): Standard wiener process
𝐷 𝑡 : Predicted demand rate at time 𝑡 (𝑑𝐷/𝑑𝑡 ≤ 0)
ℎ𝑖′: Inventory cost coefficient
(0 < ℎ1′ < ℎ2
′ < ℎ3′ < 𝑏)
𝑇: The end of time
𝑏: Unsatisfied cost coefficient
𝑐: Handling cost coefficient
Supply Chain Network (SCN)
1
RS
3
Shelter
2
DC
𝑟13𝑆13
𝐼2
𝐼𝑁𝐼1
𝐷 𝑡 𝑑𝑡~𝑁 𝐷 𝑡 𝑑𝑡, 𝐷𝑆𝐷 𝑡2𝑑𝑡
Information Updating Algorithm9
■ Predicted Demand 𝐷 𝑡 follows the normal distribution.
𝐷(𝑡) 𝑑𝑡 = 𝐷 𝑡 𝑑𝑡 + 𝐷𝑆𝐷 𝑡 𝑑𝑧(𝑡)
■ 𝐷 𝑡 is updated to subjective demand 𝐷𝑙(𝑡) based on information 𝐷 𝑡 , applying the Bayesian Estimation at an interval 𝑘𝑙(𝑘1 ≥ 𝑘2).
1
RS
3
Shelter
2
DC
Push-mode support
𝐷𝐷
𝐷
1
RS
3
Shelter
2
DC
Pull-mode support
𝐷
𝐷2
𝐷
■ The number of updates 𝑛 increases as depot 𝑙 is closer to the shelter
⟹ The Bullwhip Effect (𝑉 𝐷1 𝑡 ≥ 𝑉 𝐷2 𝑡 ∵ 𝑛1 ≤ 𝑛2)
[Dynamics]
[Distribution]
𝑡 = 𝑘1
Stochastic Optimal Control Problem10
Object Function
[Net inventory holding cost]
[Inventory holding cost]
[Inventory handling cost]
[Outflows at destination 𝑗]
[Cost functions]
State Equation (Inventory Dynamics) Initial Condition
Stochastic Optimal Control Problem11
Object Function
[Net inventory holding cost]
[Inventory holding cost]
[Inventory handling cost]
[Outflows at destination 𝑗]
[Cost functions]
State Equation (Inventory Dynamics) Initial Condition
▪𝑇𝐶𝑆𝑙 𝑡 : Changes in the inventory level in node 𝑙
⁃ min𝑇𝐶𝑆𝑙 𝑡 = 0 (∴ no transshipment)
⁃ 𝑇𝐶𝑆𝑙 𝑡 ≠ 0 means that the inventory level should change (∴ handling cost).
⁃ min𝑇𝐶𝑆𝑙 𝑡 means supply constraints for node 𝑙𝑗𝑆𝑖𝑗 𝑡 − 𝑟𝑖𝑗 𝑇𝑆𝑙𝑗 𝑡
12
- Optimal Push-mode Support -
Mathematical Analysis
Assumption13
Cost functions𝑓𝐼 𝑥 = 𝑓𝐵 𝑥 = 𝑓𝑆 𝑥 = 𝑥𝛼 , 𝛼 > 1
Assumption
1. Let 𝑡 = 𝑇 be the time when demand becomes 0, 𝐷𝑙 𝑇 = 0.
2. Demand decreases constantly over time, 𝑑𝐷𝑙 𝑡 /𝑑𝑡 = 𝐷𝑙 < 0.
3. The inventory holding cost at the shelter is twice that at the RS, ℎ𝐼𝑁′ = 1/2ℎ1
′ .
4. The DC is not ready after a disaster, 𝑆23 𝑡 = 0 ∀𝑡 ∈ 0, 𝑟12 .
The following mathematical properties will be proved:
Lemma 1. There is no need to pre-store at DC and to add stock
after a disaster, 𝐼∗ 𝑡 = 0.
Lemma 2. "Direct Supply" is effective, 𝐸 𝑆12∗ 𝑡 < 𝐸 𝑆13
∗ 𝑡 .
Theorem 1. DC is unnecessary for sequence control strategy.
Theorem 2. Maintaining inventory at the shelter is the optimal
when the penalty cost is sufficiently high, 𝐼𝑁∗ 𝑡 > 0 𝑖𝑓 𝑏 → ∞.
Optimal Control : 𝑰 𝒕 14
Optimal Inventory
■ 𝐼 𝑡 asymptotically approaches 0 from initial value 𝐼 (0) > 0.
→ Solve 𝐼∗(0) = argmin𝐼 0 𝑉∗
> 0, therefore 𝐼∗ 0 = 0
𝑉∗ = 𝑉(𝐼∗, 𝐼𝑁∗, 𝑆∗)
Pre-stock
There is no need to pre-store at DC and to add stock, 𝐼∗ 𝑡 = 0.
Lemma 1
■ Positive inventory level
■ Decreasing function
■ The limit value is 0
*
> 𝟎
Optimal Control : 𝑺𝟏𝒋 𝒕15
Optimal supply rate from RSS13∗ 𝑡S12
∗ 𝑡
Lemma 2
"Direct Supply" is effective, 𝐸 𝑆12∗ 𝑡 < 𝐸 𝑆13
∗ 𝑡 .
𝑡 ∈ 𝑟123 − 𝑟13, 𝑇 − 𝑟123
*
*𝑟13 < 𝑟123
Differentiate 𝐸 𝑆 𝑡 as follows:
We obtain,
< 0
When 𝑑𝐸 𝑆 𝑡 /𝑑𝑡 < 0, we obtain 𝐸 𝑆12∗ 𝑡 < 𝐸 𝑆13
∗ 𝑡 (∵ 𝑟13 < 𝑟123)
Optimal Control : 𝑺𝟏𝒋 𝒕16
Optimal supply rate from RSS13∗ 𝑡S12
∗ 𝑡
Lemma 2
"Direct Supply" is effective, 𝐸 𝑆12∗ 𝑡 < 𝐸 𝑆13
∗ 𝑡 .
𝑡 ∈ 𝑟123 − 𝑟13, 𝑇 − 𝑟123
*
*𝑟13 < 𝑟123
Differentiate 𝐸 𝑆 𝑡 as follows:
We obtain,
< 0
When 𝑑𝐸 𝑆 𝑡 /𝑑𝑡 < 0, we obtain 𝐸 𝑆12∗ 𝑡 < 𝐸 𝑆13
∗ 𝑡 (∵ 𝑟13 < 𝑟123)
Lemma 2
"Direct Supply" is effective, 𝐸 𝑆12∗ 𝑡 < 𝐸 𝑆13
∗ 𝑡 .
There is no need to pre-store at DC and to add stock, 𝐼∗ 𝑡 = 0.
Lemma 1
DC is unnecessary for sequence control strategy.
Theorem 1
[Dynamics]
[Optimal]
Optimal Control : 𝑰𝑵 𝒕 17
Optimal net inventory
0
𝐼𝑁 𝑡
𝑡 𝑡 + Δ𝑡
■ Dynamics of 𝐼𝑁 𝑡 is
𝐼𝑁 𝑡
𝑡 𝑡 + Δ𝑡
0
𝑏 → ∞
Maintaining inventory at the shelter is
the optimal when the penalty cost is
sufficiently high, 𝐼𝑁∗ 𝑡 > 0 𝑖𝑓 𝑏 → ∞.
Theorem 2
■ Assume 𝑏 → ∞
■ Long-term expected value is 0
■ Regression speed < 0
18
- Optimal Pull-mode Support -
Numerical Analysis
Settings19
Parameter setting
▪Prediction demand 𝐷 𝑡 = −0.15 𝑡 − 𝑇 .
▪The number of simulations is 5,000.
▪case-a「𝑅 𝑧1 𝑡 , 𝑧2 𝑡 = 1」，case-b「𝑅 𝑧1 𝑡 , 𝑧2 𝑡 ≠ 1 」
RS and DC share prediction errors not share
▪Comparing the objective functions of the push 𝑉𝑝𝑢𝑠ℎ and
pull 𝑉𝑝𝑢𝑙𝑙 𝑘1, 𝑘2 with the Monte Carlo simulation.
Analysis method
Parameter setting
Results : Comparing Push and Pull20
The sufficient conditions for control strategy change (push-mode to pull-mode) is
𝑘1 = 𝑘2, that is the centralized information system is restored.
Under the decentralized information system, pull-mode may not be effective.
𝑅 𝑧1 𝑡 , 𝑧2 𝑡 = 1(case−a)
21
Conclusion
Conclusion22
■ Dynamic optimization approach analyzed the mathematical
properties of the optimal control strategy.
⁃ Push-mode: direct supply from the RS to the shelter is optimal
transportation strategy. DC should not be used.
⁃ Pull-mode: the sufficient condition for pull-mode to be optimal
is to restore the centralized information system. Otherwise,
pull-mode may make a not always good result.
Summary
Future work
■ General network analysis (e.g., many-to-many network) can
consider significant elements such as the single point of failure.
■ Considering optimal day-to-day recovery dynamics.
Reference23
Cabinet Office in Japan (2017). White paper on disaster management 2017.
http://www.bousai.go.jp/kyoiku/panf/pdf/WP2017_DM_Full_Version.pdf. [accessed December. 14, 2017].
Cabinet Office in Japan (2018). White paper on disaster management 2018.
http://www.bousai.go.jp/kaigirep/hakusho/pdf/H30_hakusho_english.pdf. [accessed November. 25, 2018].
Chen, F., Drezner, Z., Ryan, J.K., and Simchi-Levi, D. (2000). Quantifying the bullwhip effect in a simple supply
chain:
The impact of forecasting, lead times, and information. Management science, 46(3), 436-443.
Chomilier, B. (2010). WFP logistics. World Food Programme.
Holguin-Veras, J., Taniguchi, E., Jaller, M., Aros-Vera, F., Ferreira, F., and Thompson, R.G. (2014). The tohoku
disasters: Chief lessons concerning the post disaster humanitarian logistics response and policy implications.
Transportation research part A, 69, 86-104.
Kawase, R., Urata, J., and Iryo, T. (2018). Optimal inventory distribution strategy for relief supply considering
information uncertainty after a major disaster. Informs Annual Meeting.
Mangasarian, O.L. (1966). Suffcient conditions for the optimal control of nonlinear systems. SIAM Journal on
control, 4(1), 139-152.
Maruyama, G. (1955). Continuous markov processes and stochastic equations. Rendiconti del Circolo
Matematico
di Palermo, 4(1), 48-90.
Meng, Q. and Shen, Y. (2010). Optimal control of mean-field jump-diffusion systems with delay: A stochastic
maximum principle approach. Automatica, 46(6), 1074-1080.
Sheu, J. B. (2007). An emergency logistics distribution approach for quick response to urgent relief demand in
disasters. Transportation Research Part E: Logistics and Transportation Review, 43(6), 687-709.
The Committee of Infrastructure Planning and Management (2016). Investigation report on kumamoto
earthquake: Challenges of logistics. (In Japanese).
Uhlenbeck, G.E. and Ornstein, L.S. (1930). On the theory of the brownian motion. Physical review, 36(5), 823-
841.
