![Page 1: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/1.jpg)
Dynamic LP Games with Risk-Averse Players
Alejandro Toriello
Stewart School of Industrial and Systems EngineeringGeorgia Institute of Technology
IDS Seminar, College of Business AdministrationUniversity of Illinois, Chicago
September 15, 2017
joint with Nelson Uhan (USNA)
![Page 2: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/2.jpg)
Sharing and Risk Aversion Don’t Mix:Dynamic LP Games with Risk-Averse Players
Alejandro Toriello
Stewart School of Industrial and Systems EngineeringGeorgia Institute of Technology
IDS Seminar, College of Business AdministrationUniversity of Illinois, Chicago
September 15, 2017
joint with Nelson Uhan (USNA)
![Page 3: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/3.jpg)
MotivationCalifornia cut flower industry (Nguyen/Toriello/Dessouky/Moore 13)
I Study of potential consolidation center estimated up to 1/3reduction in transport costs. Issue deemed existential by somein industry.
I Results presented before U.S. Congress.
I Application submitted to USDOT for discretionary grant.
I However, many farmers skeptical of consolidation and wary ofpotential risks.
I Looming question: Can many disparate agents agree oncooperation and cost sharing? How does aversion to riskaffect the ability to cooperate?
![Page 4: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/4.jpg)
MotivationCalifornia cut flower industry (Nguyen/Toriello/Dessouky/Moore 13)
I Study of potential consolidation center estimated up to 1/3reduction in transport costs. Issue deemed existential by somein industry.
I Results presented before U.S. Congress.
I Application submitted to USDOT for discretionary grant.
I However, many farmers skeptical of consolidation and wary ofpotential risks.
I Looming question: Can many disparate agents agree oncooperation and cost sharing? How does aversion to riskaffect the ability to cooperate?
![Page 5: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/5.jpg)
Outline
IntroductionA Crash Course in Cooperative GamesPrimer on Coherent Risk Measures
Model FormulationThe Risk-Averse Strong Sequential CoreImplications on CooperationApplication: Risk-Averse Newsvendor Games
Parting Thoughts
![Page 6: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/6.jpg)
Cost Allocation in OR Models(Static) cooperative games and the core
I Players: N = {1, . . . , n}, e.g. retailers or producers.
I Cost function: f : 2N → R, e.g. joint venture cost when somesubset of players cooperates.
I Goal: Assuming all players in N cooperate, find a costallocation χ ∈ RN that splits cost “fairly.”
I Usually models situations in which agents enter into bindingagreement.
I Especially when cooperation occurs over time.
![Page 7: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/7.jpg)
Cost Allocation in OR Models(Static) cooperative games and the core
I Players: N = {1, . . . , n}, e.g. retailers or producers.
I Cost function: f : 2N → R, e.g. joint venture cost when somesubset of players cooperates.
I Goal: Assuming all players in N cooperate, find a costallocation χ ∈ RN that splits cost “fairly.”
I Usually models situations in which agents enter into bindingagreement.
I Especially when cooperation occurs over time.
![Page 8: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/8.jpg)
Cost Allocation in OR ModelsCooperative games and the core
I Core (Gillies 59): Set of allocations χ that are
efficient:∑
N χi ≥ f(N),
stable: for U ⊆ N , no allocation ξ has∑U ξi ≥ f(U) and ξi < χi, i ∈ U.
No player or coalition can do better by defecting.
I Equivalently described by{χ ∈ RN :
∑Nχi ≥ f(N);
∑Uχi ≤ f(U), U ⊆ N
}.
I Non-empty core suggests cooperation is possible.
![Page 9: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/9.jpg)
Cost Allocation in OR ModelsCooperative games and the core
I Core (Gillies 59): Set of allocations χ that are
efficient:∑
N χi ≥ f(N),
stable: for U ⊆ N , no allocation ξ has∑U ξi ≥ f(U) and ξi < χi, i ∈ U.
No player or coalition can do better by defecting.
I Equivalently described by{χ ∈ RN :
∑Nχi ≥ f(N);
∑Uχi ≤ f(U), U ⊆ N
}.
I Non-empty core suggests cooperation is possible.
![Page 10: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/10.jpg)
Some Application ExamplesProduction, inventory and supply chain management
I Newsvendor and inventory centralization.I Chen (09), Chen/Zhang (09), Hartman/Dror (96,05),
Hartman/Dror/Shaked (00), Montrucchio/Scarsini (07),
Muller/Scarsini/Shaked (02), Ozen/Fransoo/Norde/Slikker (08),
Slikker/Fransoo/Wouters (05)
I Economic lot-sizingI Chen/Zhang (16), Gopaladesikan/Uhan/Zou (12), Toriello/Uhan
(14), van den Heuvel/Borm/Hamers (07)
I Inventory routingI Ozener/Ergun/Savelsbergh (13)
I Joint replenishmentI He/Zhang/Zhang (12), Zhang (09)
![Page 11: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/11.jpg)
Linear Production GamesOwen (75)
f(U) = minx≥0
cx
= maxλ
λ(∑
Udi)
s.t. Ax =∑
U di
s.t. λA ≤ c
TheoremIf λ is dual optimal for f(N),
χi = λdi, i ∈ N
is in the core of f .
Proof.strong duality⇒ efficiency weak duality⇒ stability
![Page 12: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/12.jpg)
Linear Production GamesOwen (75)
f(U) = minx≥0
cx
= maxλ
λ(∑
Udi)
s.t. Ax =∑
U di
s.t. λA ≤ c
TheoremIf λ is dual optimal for f(N),
χi = λdi, i ∈ N
is in the core of f .
Proof.strong duality⇒ efficiency weak duality⇒ stability
![Page 13: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/13.jpg)
Linear Production GamesOwen (75)
f(U) = minx≥0
cx = maxλ
λ(∑
Udi)
s.t. Ax =∑
U di s.t. λA ≤ c
TheoremIf λ is dual optimal for f(N),
χi = λdi, i ∈ N
is in the core of f .
Proof.strong duality⇒ efficiency weak duality⇒ stability
![Page 14: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/14.jpg)
Cooperation under UncertaintyDynamic LP games
I Players face two-stage decision process defined by
f1(U) := minx,s≥0
c1x1 + h1s1 + E[cτxτ + hτsτ ]
s.t. A1x1 − C1s1 =∑
Ud1i (stage 1)
Atxt +Bts1 − Ctst =∑
Udti, t ∈ D, (stage 2)
x’s are actions (e.g. orders) and s’s are states (e.g. inventory).
I Static core does not capture timing, need dynamic allocation.I χt
i: What player i pays if t realizes.
![Page 15: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/15.jpg)
Cooperation under UncertaintyDynamic LP games
I Players face two-stage decision process defined by
f1(U) := minx,s≥0
c1x1 + h1s1 + E[cτxτ + hτsτ ]
s.t. A1x1 − C1s1 =∑
Ud1i (stage 1)
Atxt +Bts1 − Ctst =∑
Udti, t ∈ D, (stage 2)
x’s are actions (e.g. orders) and s’s are states (e.g. inventory).
I Static core does not capture timing, need dynamic allocation.I χt
i: What player i pays if t realizes.
![Page 16: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/16.jpg)
Cooperation under UncertaintyDynamic LP games
I Suppose players implement optimal solution x, s.
I Strong sequential core (Predtetchinski/Herings/Peters 02):Allocations χ that are
stage-wise efficient: pay as you go,∑Nχ
ti ≥ ctxt + htst, t ∈ D ∪ 1,
time-consistently stable: never vulnerable to any U defecting,∑U
(χ1i + E[χτi ]
)≤ f1(U) (stage 1)∑
Uχti ≤ f
t(U, s1U ), t ∈ D (stage 2)
where f t is static problem faced by U in scenario t.
![Page 17: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/17.jpg)
Cooperation under UncertaintyDynamic LP games
I Suppose players implement optimal solution x, s.
I Strong sequential core (Predtetchinski/Herings/Peters 02):Allocations χ that are
stage-wise efficient: pay as you go,∑Nχ
ti ≥ ctxt + htst, t ∈ D ∪ 1,
time-consistently stable: never vulnerable to any U defecting,∑U
(χ1i + E[χτi ]
)≤ f1(U) (stage 1)∑
Uχti ≤ f
t(U, s1U ), t ∈ D (stage 2)
where f t is static problem faced by U in scenario t.
![Page 18: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/18.jpg)
Cooperation under UncertaintyDynamic LP games
I Suppose players implement optimal solution x, s.
I Strong sequential core (Predtetchinski/Herings/Peters 02):Allocations χ that are
stage-wise efficient: pay as you go,∑Nχ
ti ≥ ctxt + htst, t ∈ D ∪ 1,
time-consistently stable: never vulnerable to any U defecting,∑U
(χ1i + E[χτi ]
)≤ f1(U) (stage 1)∑
Uχti ≤ f
t(U, s1U ), t ∈ D (stage 2)
where f t is static problem faced by U in scenario t.
![Page 19: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/19.jpg)
Cooperation under UncertaintyDynamic LP games
Theorem (e.g. Xu/Veinott 13)
A dual-based dynamic allocation similar to Owen’s is in the strongsequential core.
I Similar results extend to multi-stage models.
![Page 20: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/20.jpg)
Example: Newsvendor GamesInventory pooling
I Each player faces uncertain demand dti, can order now,backlog or salvage later:
f1nv(U) := minx,s≥0
x1 + E[bxτ − vsτ ]
s.t. x1 + xt − st =∑
Udti, t ∈ D.
I Suppose scenarios ordered by∑
N dti. Then critical index is
t := maxt∈D
P(demand ≥∑
Ndti) ≥ 1−v
b−v .
Optimal order is x1 =∑
N dti.
![Page 21: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/21.jpg)
Example: Newsvendor GamesInventory pooling
I Each player faces uncertain demand dti, can order now,backlog or salvage later:
f1nv(U) := minx,s≥0
x1 + E[bxτ − vsτ ]
s.t. x1 + xt − st =∑
Udti, t ∈ D.
I Suppose scenarios ordered by∑
N dti. Then critical index is
t := maxt∈D
P(demand ≥∑
Ndti) ≥ 1−v
b−v .
Optimal order is x1 =∑
N dti.
![Page 22: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/22.jpg)
Example: Newsvendor GamesInventory pooling
I If s1i = dti, i.e. players split order according to demand in t,
χ1i = dti, χti =
b(dti − dti), t > t
−v(dti − dti), t < t
0, t = t
is in the SSC.
I Each player pays to order their demand under scenario t.
I Second-stage order/salvage and side payments occur at priceb/v depending on realization.
![Page 23: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/23.jpg)
Coherent Risk MeasuresArtzner/Delbaen/Eber/Heath (99)
I We now suppose players may be risk-averse.
I Assume risk preferences captured by coherent risk measure,ρ : RD → R with
monotonicity: ρ(χτ ) ≤ ρ(ξτ ) if χ ≤ ξ,
translation invariance: ρ(χτ + κ) = ρ(χτ ) + κ for constant κ,
positive homogeneity: ρ(κχτ ) = κρ(χτ ) for constant κ ≥ 0,
subadditivity: ρ(χτ + ξτ ) ≤ ρ(χτ ) + ρ(ξτ ) (convexity)
I Expectation is risk-neutral, additive instead of subadditive.
![Page 24: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/24.jpg)
Coherent Risk MeasuresArtzner/Delbaen/Eber/Heath (99)
I We now suppose players may be risk-averse.
I Assume risk preferences captured by coherent risk measure,ρ : RD → R with
monotonicity: ρ(χτ ) ≤ ρ(ξτ ) if χ ≤ ξ,
translation invariance: ρ(χτ + κ) = ρ(χτ ) + κ for constant κ,
positive homogeneity: ρ(κχτ ) = κρ(χτ ) for constant κ ≥ 0,
subadditivity: ρ(χτ + ξτ ) ≤ ρ(χτ ) + ρ(ξτ ) (convexity)
I Expectation is risk-neutral, additive instead of subadditive.
![Page 25: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/25.jpg)
Coherent Risk MeasuresArtzner/Delbaen/Eber/Heath (99)
I We now suppose players may be risk-averse.
I Assume risk preferences captured by coherent risk measure,ρ : RD → R with
monotonicity: ρ(χτ ) ≤ ρ(ξτ ) if χ ≤ ξ,
translation invariance: ρ(χτ + κ) = ρ(χτ ) + κ for constant κ,
positive homogeneity: ρ(κχτ ) = κρ(χτ ) for constant κ ≥ 0,
subadditivity: ρ(χτ + ξτ ) ≤ ρ(χτ ) + ρ(ξτ ) (convexity)
I Expectation is risk-neutral, additive instead of subadditive.
![Page 26: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/26.jpg)
Coherent Risk MeasuresArtzner/Delbaen/Eber/Heath (99)
I We now suppose players may be risk-averse.
I Assume risk preferences captured by coherent risk measure,ρ : RD → R with
monotonicity: ρ(χτ ) ≤ ρ(ξτ ) if χ ≤ ξ,
translation invariance: ρ(χτ + κ) = ρ(χτ ) + κ for constant κ,
positive homogeneity: ρ(κχτ ) = κρ(χτ ) for constant κ ≥ 0,
subadditivity: ρ(χτ + ξτ ) ≤ ρ(χτ ) + ρ(ξτ ) (convexity)
I Expectation is risk-neutral, additive instead of subadditive.
![Page 27: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/27.jpg)
Coherent Risk MeasuresArtzner/Delbaen/Eber/Heath (99)
I We now suppose players may be risk-averse.
I Assume risk preferences captured by coherent risk measure,ρ : RD → R with
monotonicity: ρ(χτ ) ≤ ρ(ξτ ) if χ ≤ ξ,
translation invariance: ρ(χτ + κ) = ρ(χτ ) + κ for constant κ,
positive homogeneity: ρ(κχτ ) = κρ(χτ ) for constant κ ≥ 0,
subadditivity: ρ(χτ + ξτ ) ≤ ρ(χτ ) + ρ(ξτ ) (convexity)
I Expectation is risk-neutral, additive instead of subadditive.
![Page 28: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/28.jpg)
Coherent Risk MeasuresArtzner/Delbaen/Eber/Heath (99)
TheoremAny coherent risk measure has a robust representation as aworst-case expectation over a closed, convex set of distributions,
ρ(χτ ) = maxq∈Q
Eq[χτ ]
for Q ⊆ ∆D.
I Our results extend to multi-stage models using conditionalrisk mappings (Ruszczynski/Shapiro 06).
![Page 29: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/29.jpg)
Including Risk Aversion in LP Game
I Suppose players face same situation.
I Linear dynamics, additive demand/requirements, two stages.
I But each player i assesses uncertain costs based on coherentrisk measure ρi.
I What solution should they implement?
I Given a solution, how are costs split? What do players think isfair?
![Page 30: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/30.jpg)
The Risk-Averse Strong Sequential Core
If the players implement solution x, s, the risk-averse SSC hasallocations χ with
stage-wise efficiency:∑Nχ
ti ≥ ctxt + htst, t ∈ D ∪ 1,
time-consistent stability: for coalition U ⊆ N ,
stage 1 No other solution x′, s′ satisfies U ’s demandand has a stage-wise efficient allocation ξ with
ξ1i + ρi(ξτi ) < χ1
i + ρi(χτi ), i ∈ U,
stage 2∑
Uχti ≤ f t(U, s1U ), t ∈ D.
![Page 31: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/31.jpg)
The Risk-Averse Strong Sequential Core
If the players implement solution x, s, the risk-averse SSC hasallocations χ with
stage-wise efficiency:∑Nχ
ti ≥ ctxt + htst, t ∈ D ∪ 1,
time-consistent stability: for coalition U ⊆ N ,
stage 1 No other solution x′, s′ satisfies U ’s demandand has a stage-wise efficient allocation ξ with
ξ1i + ρi(ξτi ) < χ1
i + ρi(χτi ), i ∈ U,
stage 2∑
Uχti ≤ f t(U, s1U ), t ∈ D.
![Page 32: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/32.jpg)
The Risk-Averse Strong Sequential Core
If the players implement solution x, s, the risk-averse SSC hasallocations χ with
stage-wise efficiency:∑Nχ
ti ≥ ctxt + htst, t ∈ D ∪ 1,
time-consistent stability: for coalition U ⊆ N ,
stage 1 No other solution x′, s′ satisfies U ’s demandand has a stage-wise efficient allocation ξ with
ξ1i + ρi(ξτi ) < χ1
i + ρi(χτi ), i ∈ U,
stage 2∑
Uχti ≤ f t(U, s1U ), t ∈ D.
![Page 33: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/33.jpg)
The Risk-Averse Strong Sequential Core
TheoremThe SSC for solution x, s is the set of allocations χ with∑
Nχti ≥ ctxt + htst, t ∈ D ∪ 1, (efficiency)∑
U
(χ1i + ρi(χ
τi ))≤ f1(U), U ⊆ N (stage-1 stable)∑
Uχti ≤ f t(U, s1U ), U ⊆ N, t ∈ D, (stage-2 stable)
where
f1(U) := minx,s≥0;ξ
∑U
(ξ1i + ρi(ξ
τi ))
s.t. A1x1 − C1s1 =∑
Ud1i
Atxt +Bts1 − Ctst =∑
Udti, t ∈ D,∑
Uξti ≥ ctxt + htst, t ∈ D ∪ 1.
![Page 34: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/34.jpg)
The Risk-Averse Strong Sequential Core
TheoremThe SSC for solution x, s is the set of allocations χ with∑
Nχti ≥ ctxt + htst, t ∈ D ∪ 1, (efficiency)∑
U
(χ1i + ρi(χ
τi ))≤ f1(U), U ⊆ N (stage-1 stable)∑
Uχti ≤ f t(U, s1U ), U ⊆ N, t ∈ D, (stage-2 stable)
where
f1(U) := minx,s≥0;ξ
∑U
(ξ1i + ρi(ξ
τi ))
s.t. A1x1 − C1s1 =∑
Ud1i
Atxt +Bts1 − Ctst =∑
Udti, t ∈ D,∑
Uξti ≥ ctxt + htst, t ∈ D ∪ 1.
![Page 35: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/35.jpg)
The Risk-Averse Strong Sequential CoreStructural implications
If the players implement x, s and use allocation χ in the SSC,
risk optimality: x, s and χ are optimal for f1(N),∑N
(χ1i + ρi(χ
τi ))
= f1(N),
convexity: if the risk measures ρi are convex, so is the SSC.
![Page 36: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/36.jpg)
Consequences for CooperationRisk alignment
Corollary
If χ is in the SSC, ∑Nρi(χ
τi ) =
∑NEq[χτi ],
for some q ∈ ∆D.
I The allocation must be risk-aligned; the same distributionmust yield worst-case expected cost for all players.
I E.g. player A cannot hope for low demand if player B hopesfor high demand.
![Page 37: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/37.jpg)
Consequences for CooperationRisk alignment
Corollary
If χ is in the SSC, ∑Nρi(χ
τi ) =
∑NEq[χτi ],
for some q ∈ ∆D.
I The allocation must be risk-aligned; the same distributionmust yield worst-case expected cost for all players.
I E.g. player A cannot hope for low demand if player B hopesfor high demand.
![Page 38: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/38.jpg)
Consequences for CooperationPooling versus stability
Corollary
If player j is equally or less risk-averse than all others,
Qj ⊆ Qi, i ∈ N \ j,
then f1 can be optimized by assigning all cost to j.
I Optimization requires the pooling of cost (subadditivity),stability requires this cost spread out without increasing risk.
I May only be possible by assigning all uncertainty to j, whileother players get scenario-independent side payments.
![Page 39: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/39.jpg)
Consequences for CooperationPooling versus stability
Corollary
If player j is equally or less risk-averse than all others,
Qj ⊆ Qi, i ∈ N \ j,
then f1 can be optimized by assigning all cost to j.
I Optimization requires the pooling of cost (subadditivity),stability requires this cost spread out without increasing risk.
I May only be possible by assigning all uncertainty to j, whileother players get scenario-independent side payments.
![Page 40: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/40.jpg)
Consequences for CooperationPooling versus stability
Corollary
If player j is equally or less risk-averse than all others,
Qj ⊆ Qi, i ∈ N \ j,
then f1 can be optimized by assigning all cost to j.
I Optimization requires the pooling of cost (subadditivity),stability requires this cost spread out without increasing risk.
I May only be possible by assigning all uncertainty to j, whileother players get scenario-independent side payments.
![Page 41: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/41.jpg)
Consequences for CooperationCompatible beliefs
Corollary
The SSC is empty if ⋂NQi = ∅.
I If players cannot agree on a common belief (distribution)about the future, they cannot cooperate.
I E.g. player A believes tomorrow will be sunny, B believes itwill be cloudy. Each can bet arbitrary amount on their beliefto create artificial risk arbitrage.
![Page 42: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/42.jpg)
Consequences for CooperationCompatible beliefs
Corollary
The SSC is empty if ⋂NQi = ∅.
I If players cannot agree on a common belief (distribution)about the future, they cannot cooperate.
I E.g. player A believes tomorrow will be sunny, B believes itwill be cloudy. Each can bet arbitrary amount on their beliefto create artificial risk arbitrage.
![Page 43: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/43.jpg)
Risk-Averse Newsvendor Games
I Assume players assess risk with same comonotonic measure, ρ.
f1nv(U) := minx,s≥0
x1 + ρ(bxτ − vsτ )
s.t. x1 + xt − st =∑
Udti, t ∈ D.
comonotonicity: ρ(χτ + ξτ ) = ρ(χτ ) + ρ(ξτ ) if χ, ξ comonotonic
I Critical scenario index t defined with worst-case distribution.
I Still optimal to order x1 =∑
N dti in first stage.
![Page 44: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/44.jpg)
Risk-Averse Newsvendor Games
I Assume players assess risk with same comonotonic measure, ρ.
f1nv(U) := minx,s≥0
x1 + ρ(bxτ − vsτ )
s.t. x1 + xt − st =∑
Udti, t ∈ D.
comonotonicity: ρ(χτ + ξτ ) = ρ(χτ ) + ρ(ξτ ) if χ, ξ comonotonic
I Critical scenario index t defined with worst-case distribution.
I Still optimal to order x1 =∑
N dti in first stage.
![Page 45: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/45.jpg)
Risk-Averse Newsvendor Games(Not so) good news
I Recall solution and allocation with s1i = dti and
χ1i = dti, χti =
b(dti − dti), t > t
−v(dti − dti), t < t
0, t = t.
TheoremSuppose players’ demand is comonotonic, d1i ≤ d2i ≤ . . . for i ∈ N .Then χ is in the SSC.
I If demand is comonotonic, players perceive least benefit fromcooperation.
I χ mirrors what each player would incur on their own.
![Page 46: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/46.jpg)
Risk-Averse Newsvendor Games(Not so) good news
I Recall solution and allocation with s1i = dti and
χ1i = dti, χti =
b(dti − dti), t > t
−v(dti − dti), t < t
0, t = t.
TheoremSuppose players’ demand is comonotonic, d1i ≤ d2i ≤ . . . for i ∈ N .Then χ is in the SSC.
I If demand is comonotonic, players perceive least benefit fromcooperation.
I χ mirrors what each player would incur on their own.
![Page 47: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/47.jpg)
Risk-Averse Newsvendor GamesBad news: an example with empty SSC
I Two players, exactly complementary demand,
d011 = d102 = 0, d101 = d012 = 1.
Ideal situation: Optimal joint order is 1, no uncertainty.
I Risk measure defined by Q = {(q, 1− q), (1− q, q)}.
I If v > 0, b big enough, q 6= 1/2, the SSC is empty.
I No matter how players split the order, the SSC requires the“losing” player to receive v per unit, but the “winning” playerto not pay anything.
I Holds for any q 6= 1/2, any risk measure that isn’t risk-neutral.
![Page 48: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/48.jpg)
Risk-Averse Newsvendor GamesBad news: an example with empty SSC
I Two players, exactly complementary demand,
d011 = d102 = 0, d101 = d012 = 1.
Ideal situation: Optimal joint order is 1, no uncertainty.
I Risk measure defined by Q = {(q, 1− q), (1− q, q)}.
I If v > 0, b big enough, q 6= 1/2, the SSC is empty.
I No matter how players split the order, the SSC requires the“losing” player to receive v per unit, but the “winning” playerto not pay anything.
I Holds for any q 6= 1/2, any risk measure that isn’t risk-neutral.
![Page 49: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/49.jpg)
Risk-Averse Newsvendor GamesBad news: an example with empty SSC
I Two players, exactly complementary demand,
d011 = d102 = 0, d101 = d012 = 1.
Ideal situation: Optimal joint order is 1, no uncertainty.
I Risk measure defined by Q = {(q, 1− q), (1− q, q)}.
I If v > 0, b big enough, q 6= 1/2, the SSC is empty.
I No matter how players split the order, the SSC requires the“losing” player to receive v per unit, but the “winning” playerto not pay anything.
I Holds for any q 6= 1/2, any risk measure that isn’t risk-neutral.
![Page 50: 0.7em Dynamic LP Games with Risk-Averse Playersatoriello3/rav_lp_game_pres_uic_2017-09-15.pdf · Dynamic LP Games with Risk-Averse Players Alejandro Toriello Stewart School of Industrial](https://reader030.vdocuments.site/reader030/viewer/2022040816/5e5f5b711168771f141210ef/html5/thumbnails/50.jpg)
Conclusions
I In dynamic LP games, cooperation always possible whenplayers are risk neutral.
I When players are risk-averse, cooperation much more difficult.
I SSC can easily be empty, implying cooperation unlikely.
I We can only guarantee SSC is non-empty when cooperation isleast beneficial.
I Results suggest risk neutrality is important necessarycondition in cooperation.
I Risk aversion may explain lack of cooperation in situationswhere risk-neutral models predict it.