Finding a Maximum Stable Flow in a Multi-agent ... - LAAS-CNRS

0 downloads 0 Views 885KB Size Report
Feb 28, 2014 - maximize its profit. Multi-objective optimization : Pareto optima. Stability. Given a strategy, no agent should be able to increase its profit by itself, ...
Finding a Maximum Stable Flow in a Multi-agent Network with Controllable Capacities Nadia CHAABANE(1)

Cyril Briand (1)

(1) LAAS-CNRS,

Marie-Jose´ Huguet(1)

University of Toulouse

ROADEF’14 - 26-28 February 2014

1 / 37

Outline

1

Introduction

2

Multi-agent Network with Controllable Capacities Problem statement Example

3

Finding a maximum-flow stable strategy Characterization of Nash equilibria Problem complexity MILP formulation

4

Conclusion

2 / 37

Outline

1

Introduction

2

Multi-agent Network with Controllable Capacities Problem statement Example

3

Finding a maximum-flow stable strategy Characterization of Nash equilibria Problem complexity MILP formulation

4

Conclusion

3 / 37

Introduction Optimization in a multi-agent context A set of agents is involved in a common decision-making process. Every agent controls its own decision variables (strategy) and is interested in maximizing its profit. The agents’ profit can depend on the strategies of the other agents. ⇒ Which strategy to adopt in order to fulfill the social goal, provided agents’ objectives are satisfied ?

4 / 37

Introduction Optimization in a multi-agent context A set of agents is involved in a common decision-making process. Every agent controls its own decision variables (strategy) and is interested in maximizing its profit. The agents’ profit can depend on the strategies of the other agents. ⇒ Which strategy to adopt in order to fulfill the social goal, provided agents’ objectives are satisfied ?

Multi-objective Optimization opt F (x) = (Z1 (x), Z2 (x), . . . , Zm (x)) s.c. x ∈ ω A = {A1 , . . . , Am } Agents set x = (x1 , . . . , xm ) Vector of agents’ strategies. Zu (x) profit of agent Au 5 / 37

Strategy requirement Efficiency Only non-dominated strategies are of interest because every agent wants to maximize its profit. Multi-objective optimization : Pareto optima

Stability Given a strategy, no agent should be able to increase its profit by itself, while decreasing the profit of others. Game theory : Nash equilibria.

Efficiency Vs. Stability Price of cooperation I I

Price of anarchy : PA=(GOF value in the worst NE)/(OPT) a Price of stability : PS=(GOF value in the best NE)/(OPT)

a. Global Objective Function (GOF). 6 / 37

Outline

1

Introduction

2

Multi-agent Network with Controllable Capacities Problem statement Example

3

Finding a maximum-flow stable strategy Characterization of Nash equilibria Problem complexity MILP formulation

4

Conclusion

7 / 37

Multi-agent Network with Controllable Capacities Definition Multi-agent Network < G, A, Q, Q, C, π, W > : G = (V , E) is a Network : set of nodes V where s, t ∈ V are the source and the sink nodes ; set of arcs E each one having its controllable capacity and receiving a flow.

A = {A1 , . . . , Au , . . . , Am } a set of m agents ; Each arc (i, j) belongs to exactly one agent ; Eu set of arcs belonging to agent Au ;

Q = {q i,j }(i,j)∈E et Q = {q i,j }(i,j)∈E referred as the vectors of normal and maximum arc capacities. qi,j ∈ [q i,j , q i,j ] is the capacity of arc (i, j).

8 / 37

Multi-agent Network with Controllable Capacities Definition (2) C = {ci,j } is the vector of costs. where ci,j is the unitary costs incurred by agent Au for increasing qi,j by one unit ; The Ptotal cost incurred by agent Au for increasing its arc capacities is (i,j)∈Eu ci,j (qi,j − q i,j )

π the reward given by a final customer per unit of flow circulating in the Network. W = {wu } sharing policy of rewards among the agents. The total reward for agent Au is wu × π × F

F circulating flow in the Network. fi,j flow circulating on arc (i, j) holds fi,j ≤ qi,j where qi,j ∈ [q i,j , q i,j ]

9 / 37

Multi-agent Network with Controllable Capacities Multi-objective mathematical program Max s.c. (i) (ii) (iii)

(Z1 (S), Z2 (S), . . . , Zm (S)) fP i,j ≤ qi,j , ∀(i, j) P∈ E (i,j)∈E fi,j = (j,i)∈E fji , ∀j ∈ V {s, t} q i,j ≤ qi,j ≤ q i,j , ∀(i, j) ∈ E fi,j ≥ 0, ∀(i, j) ∈ E

Profit of agent Au is P Zu (S) = wu π (F (S) − F ) − (i,j)∈Eu ci,j (qi,j − q i,j ) S = (Q1 , . . . , Qm ) is the vector of individual strategies of all agents, where : Q ≤ Qu ≤ Q is the vector of the capacities chosen by agent Au for the arcs belonging to him. 10 / 37

Multi-agent Network with Controllable Capacities

Social goal The customer wants to maximize the flow in the Network and give a reward to agents.

Strategies of the agents The capacities chosen by the agent Au for its arcs corresponds to the individual strategy of Au .

Type of the game Non-cooperative game between the agents where each of them aims to maximize its profit.

11 / 37

Example Example of MA-MCF Network Network G(V , U) with controllable capacities Two agents : Blue (AB ) and Green (AG ) Reward and sharing policy : π = 120 and wB = wG =

1 2

F IGURE: Example of MA-MCF Network with two agents

12 / 37

Example (2) Increasing the flow Find an increasing path P in the residual graph such that costu (P) < wu × π for all Au , with : P P costu (P) = (i,j)∈P + ∩Eu ci,j − (i,j)∈P − ∩Eu ci,j

F IGURE: Example of MA-MCF Network with two agents

13 / 37

Example (3) Decreasing the flow Find a decreasing path P in the residual graph such that profitu (P) > wu × π for all Au , with : P P profitu (P) = (i,j)∈P − ∩E ci,j − (i,j)∈P + ∩E ci,j u

u

F IGURE: Example of MA-MCF Network with two agents 14 / 37

Example (4) Efficiency Vs. Stability

Strategy S0

Flow 0

Reward 0

costG 0

costB 0

ZG 0

ZB 0

F IGURE: Strategy S0

15 / 37

Example (5) Efficiency Vs. Stability

Strategy S0 S1

Flow 0 1

Reward 0 120

costG 0 30

costB 0 30

ZG 0 30

ZB 0 30

F IGURE: Strategy S1

16 / 37

Example (6) Efficiency Vs. Stability Strategy S0 S1 S2

Flow 0 1 2

Reward 0 120 240

costG 0 30 80

costB 0 30 80

ZG 0 30 40

ZB 0 30 40

F IGURE: Strategy S2 17 / 37

Example (7) Efficiency Vs. Stability

Strategy S0 S1 S2

Flow 0 1 2

Efficiency Vs. Stability

Reward 0 120 240

costG 0 30 80

costB 0 30 80

ZG 0 30 40

ZB 0 30 40

Strategy S2

S2 maximizes the flow. Is S2 Nash-stable ?

18 / 37

Example (8) Efficiency Vs. Stability

Strategy S0 S1 S2

Flow 0 1 2

Reward 0 120 240

costG 0 30 80

costB 0 30 80

ZG 0 30 40

ZB 0 30 40

F IGURE: Residual graph corresponding to the strategy S2 19 / 37

Example (9) Efficiency Vs. Stability

Strategy S0 S1 S2

Flow 0 1 2

Reward 0 120 240

costG 0 30 80

costB 0 30 80

ZG 0 30 40

ZB 0 30 40

F IGURE: Profitable decreasing path for agent AG 20 / 37

Example (10) Efficiency Vs. Stability Strategy S0 S1 S2

Flow 0 1 2

Reward 0 120 240

costG 0 30 80

costB 0 30 80

ZG 0 30 40

ZB 0 30 40

Agent AG can improve its own profit by modifying unilaterally its strategy. Strategy S20

Flow 1

Reward 120

costG 10

costB 80

ZG 50

ZB -20

21 / 37

Example (11)

Efficiency Vs. Stability

S2 is a Pareto Optimum but not a Nash equilibrium S1 is not Pareto Optimum but a Nash equilibrium

Objective Find a strategy that maximizes the flow, which is a Nash equilibrium.

22 / 37

Outline

1

Introduction

2

Multi-agent Network with Controllable Capacities Problem statement Example

3

Finding a maximum-flow stable strategy Characterization of Nash equilibria Problem complexity MILP formulation

4

Conclusion

23 / 37

Characterization of Nash equilibrium

Nash equilibrium A non-poor strategy S is a Nash equilibrium if and only if there is no profitable path P such that : I

I

∀Au ∈ A, ∀P ∈ P such that (i, j) ∈ Eu costu (P) > wu × π

(1)

profitu (P) < wu × π

(2)

∀Au ∈ A, ∀P ∈ P

24 / 37

Problem complexity Is there a Nash equilibrium strategy S such that F (S) > ϕ ? Strongly NP-complete.

Reduction from 3-partition problem Consider a set ζ = {a1 , . . . , aK } of K = 3 × k integer, such that a : each integer ai ∈]B/4, B/2], for all i = 1, . . . , K PK i=1 ai = k × B

Deciding whether ζ can be partitioned into k subsets so that the sum of integers in each subset is equal to B ? a. Garey, M.R. and Johnson, D.S. (1979). Computers and Intracability : A guide to the Theory of NP-Completeness. W.H. Freeman and Co., New York, USA.

25 / 37

Problem complexity

Proof : Reduction from a 3-partition problem Reduce an instance of the 3-partition problem to an instance of MA-MCF problem. Build up a network G from the 3-partition problem instance, such that K = 9, k = 3, B = 24 and ζ = {7, 8, 7, 7, 7, 8, 9, 10, 9}. Multi-agent Network flow with k = 3 agents each agent owns K = 9 arcs each integer ai is represented by 3 parallel arcs with capacity qi,j ∈ [0, 1] and cost ci,j = ai flow = 0 Reward of the agent Au : wu × π = B +  where  is small positive number.

26 / 37

Problem complexity Proof : Reduction from a 3-partition problem Build up a network G from the 3-partition problem instance, where K = 9, k = 3, B = 24 and ζ = {7, 8, 7, 7, 7, 8, 9, 10, 9}. Multi-agent Network flow with k = 3 agents each agent owns K = 9 arcs each integer ai is represented by 3 parallel arcs with capacity qi,j ∈ [0, 1] and cost ci,j = ai flow = 0 Reward of the agent Au : wu × π = B +  where  is small positive number.

0

(e9 , 7)

(e18 ,7)

(e2 , 7)

(e1 ,8)

(e0 , 7)

(e3 , 7)

(e5 ,8)

(e4 , 7)

(e7 ,10)

(e6 ,9)

(e8 ,9) (e17 ,9)

(e10 ,8) 1

2

(e19 ,8)

3

4



5

6

7

8

9

(e26 ,9)

A1 A2 A3

F IGURE: Reduction from a 3-partition instance problem with k = 3

27 / 37

Problem complexity Proof : Reduction from a 3-partition problem (2) Find whether it exists a Nash strategy such that the flow is strictly greater than 0? Find a profitable path such that the cost for each agent must not exceed the reward B +  = 24 + 

0

(e9 , 7)

(e18 ,7)

(e2 , 7)

(e1 ,8)

(e0 , 7)

(e3 , 7)

(e5 ,8)

(e4 , 7)

(e7 ,10)

(e6 ,9)

(e8 ,9) (e17 ,9)

(e10 ,8) 1

2

(e19 ,8)

3

4



5

6

7

8

9

(e26 ,9)

A1 A2 A3

F IGURE: Reduction from a 3-partition instance problem with k = 3

28 / 37

Problem complexity

Proof : Reduction from a 3-partition problem (2) Find, whether it exists, a Nash strategy such that the flow is strictly greater than 0? Find a profitable path such that the cost for each agent must not exceed the reward B +  = 24 + 

Problem complexity Finding a Nash Equilibrium that maximizes the flow in a multi-agent Network with controllable capacities is NP−hard.

29 / 37

Mathematical formulation

Finding a NE maximizing the flow Max s.t.

F

(1)

P

(iii)

  0 ∀i 6= s, t F ,i =s , ∀i ∈ V (i,j)∈E fi,j − (j,i)∈Er fj,i =  −F , i = t q i,j ≤ qi,j ≤ q i,j , ∀(i, j) ∈ E

(iv )

profitu (P) < wu × π, ∀P ∈ Gf

P

Stability constraints

fi,j ≥ 0, ∀(i, j) ∈ E

30 / 37

Mathematical formulation

How to formulate the stability constraints ? Linearize the Nash stability constraints as constraints of MILP. (iv ) profitu (P) < wu × π, ∀P ∈ Gf

Stability constraints

fi,j ≥ 0, ∀(i, j) ∈ E Formulating the constraints based on the identification of decreasing path P ∈ Gf having the maximum profit profitu (P).

31 / 37

MILP formulation

Constraint reformulation Constraint reformulation is based on identification of path P ∈ P(S) having maximal profitu (P). profitu (P) < wu × π, ∀P ∈ P(S) ⇒ max profitu (P) < wu × π ∀P∈P

All path have to be computed. Computing the longest path for a given strategy using primal/dual constraints.

32 / 37

MILP formulation Using primal/dual constraints : Primal : longest path Min s.t. (i) (ii)

u tn+1

tju − tiu ≥ δFi,j,u , ∀(i, j) ∈ E, ∀Au ∈ A tiu − tju ≥ δBj,i,u , ∀(i, j) ∈ E, ∀Au ∈ A tiu ≥ 0, ∀i ∈ V

Dual Max s.t.

P

(i)

P

(i,j)∈E

ϕui,j × δFi,j,u +

P

(i,j)∈Er

ϕui,j × δBj,i,u

  0 ∀i 6= s, t u u −1 , i = s , ∀i ∈ V , ∀Au ∈ A (i,j)∈E∪Er ϕi,j − (i,j)∈E∪Er ϕj,i =  1,i =t φui,j ∈ {0, 1} ∀(i, j) ∈ E, ∀Au ∈ A P

33 / 37

Outline

1

Introduction

2

Multi-agent Network with Controllable Capacities Problem statement Example

3

Finding a maximum-flow stable strategy Characterization of Nash equilibria Problem complexity MILP formulation

4

Conclusion

34 / 37

Conclusion Multi-agent min cost flow Multi-agent min cost flow problem with controllable capacities. • Efficiency and stability notions. • Optimization problem = Find a Nash equilibrium that maximizes

the network flow. NP-hard in the strong sense.

35 / 37

Conclusion Multi-agent min cost flow Multi-agent min cost flow problem with controllable capacities. • Efficiency and stability notions. • Optimization problem = Find a Nash equilibrium that maximizes

the network flow. NP-hard in the strong sense.

Perspectives • Work in progress Mixed Integer Linear Programming (MILP) in order to find the best Nash equilibrium that maximizes the flow. • Future work Distributed approach to find efficient strategies that are Nash-stable. 36 / 37

Thank you for your attention !

37 / 37

Suggest Documents