One-Way Flow Nash Networks

1. One-Way Flow Nash Networks Frank Thuijsman Joint work with Jean Derks, Jeroen Kuipers, Martijn Tennekes

2. Outline • The Model of One-Way Flow Networks • An Existence Result • A Counterexample

3. Outline • The Model of One-Way Flow Networks • An Existence Result • A Counterexample

4. Outline • The Model of One-Way Flow Networks • An Existence • A Counterexample

5. The Model of One-Way Flow Networks Network Formation Game(N,v,c)

6. The Model of One-Way Flow Networks Network Formation Game(N,v,c) • N={1,2,3,…,n}

7. The Model of One-Way Flow Networks Network Formation Game(N,v,c) • N={1,2,3,…,n} • vijis the profit for agentifor being connected to agentj

8. The Model of One-Way Flow Networks Network Formation Game(N,v,c) • N={1,2,3,…,n} • vijis the profit for agentifor being connected to agentj • cijis the cost for agentifor making a link to agentj

9. The Model of One-Way Flow Networks Network Formation Game(N,v,c) • N={1,2,3,…,n} • vijis the profit for agentifor being connected to agentj • cijis the cost for agentifor making a link to agentj Exampleof a networkg 6● ●5 1● ●4 2● 3●

10. The Model of One-Way Flow Networks Network Formation Game(N,v,c) • N={1,2,3,…,n} • vijis the profit for agentifor being connected to agentj • cijis the cost for agentifor making a link to agentj Exampleof a networkg Agent 1 is connected to agents 3,4,5 and 6 and obtains profits v13 , v14 , v15 , v16 6● ●5 1● ●4 2● 3●

11. The Model of One-Way Flow Networks Network Formation Game(N,v,c) • N={1,2,3,…,n} • vijis the profit for agentifor being connected to agentj • cijis the cost for agentifor making a link to agentj Exampleof a networkg Agent 1 is connected to agents 3,4,5 and 6 and obtains profits v13 , v14 , v15 , v16 Agent 1 is not connected to agent 2 6● ●5 1● ●4 2● 3●

12. The Model of One-Way Flow Networks Network Formation Game(N,v,c) • N={1,2,3,…,n} • vijis the profit for agentifor being connected to agentj • cijis the cost for agentifor making a link to agentj Exampleof a networkg Agent 1 is connected to agents 3,4,5 and 6 and obtains profits v13 , v14 , v15 , v16 Agent 1 is not connected to agent 2 Agent 1 has to pay c13 for the link (3,1) 6● ●5 1● ●4 2● 3●

13. The Model of One-Way Flow Networks Network Formation Game(N,v,c) • N={1,2,3,…,n} • vijis the profit for agentifor being connected to agentj • cijis the cost for agentifor making a link to agentj Exampleof a networkg The payoff π1(g) for agent 1 in network g is π1(g) = v13 + v14 + v15 + v16 - c13 6● ●5 1● ●4 2● 3●

14. The Model of One-Way Flow Networks More generally πi(g) = ∑j ЄNi(g)vij– ∑j ЄDi(g)cij where Ni(g) is the set of agents that i is connected to, and where Di(g) is the set of agents that i is directly connected to. 6● ●5 1● ●4 2● 3●

15. The Model of One-Way Flow Networks More generally πi(g) = ∑j ЄNi(g)vij– ∑j ЄDi(g)cij where Ni(g) is the set of agents that i is connected to, and where Di(g) is the set of agents that i is directly connected to. An action for agent i is any subset Sof N\{i} indicating the set of agents that i connects to directly 6● ●5 1● ●4 2● 3●

16. The Model of One-Way Flow Networks More generally πi(g) = ∑j ЄNi(g)vij– ∑j ЄDi(g)cij where Ni(g) is the set of agents that i is connected to, and where Di(g) is the set of agents that i is directly connected to. An action for agent i is any subset Sof N\{i} indicating the set of agents that i connects to directly A network g is a Nash network if each agent i is playing a best response in terms of his individual payoff πi(g) 6● ●5 1● ●4 2● 3●

17. A Closer Look at Nash Networks A network g is a Nash network if for each agent i πi(g) ≥ πi(g-i + {(j,i): jЄS}) for all subsets S of N\{i} 6● g ●5 1● ●4 2● 3●

18. A Closer Look at Nash Networks A network g is a Nash network if for each agent i πi(g) ≥ πi(g-i + {(j,i): jЄS}) for all subsets S of N\{i} Here g-i denotes the network derived from g by removing all direct links of agenti 6● g ●5 1● ●4 2● 3●

19. A Closer Look at Nash Networks A network g is a Nash network if for each agent i πi(g) ≥ πi(g-i + {(j,i): jЄS}) for all subsets S of N\{i} Here g-i denotes the network derived from g by removing all direct links of agenti 6● 6● g ●5 ●5 1● 1● ●4 ●4 2● 2● 3● 3●

20. A Closer Look at Nash Networks A network g is a Nash network if for each agent i πi(g) ≥ πi(g-i + {(j,i): jЄS}) for all subsets S of N\{i} Here g-i denotes the network derived from g by removing all direct links of agenti 6● 6● g-3 g ●5 ●5 1● 1● ●4 ●4 2● 2● 3● 3●

21. A Closer Look at Nash Networks A network g is a Nash network if for each agent i πi(g) ≥ πi(g-i + {(j,i): jЄS}) for all subsets S of N\{i} A set S that maximizes the right-hand side of above expression is called a best response for agent i to the network g

22. A Closer Look at Nash Networks A network g is a Nash network if for each agent i πi(g) ≥ πi(g-i + {(j,i): jЄS}) for all subsets S of N\{i} A set S that maximizes the right-hand side of above expression is called a best response for agent i to the network g In a Nash network all agents are linked to their best responses

23. A Closer Look at Nash Networks A network g is a Nash network if for each agent i πi(g) ≥ πi(g-i + {(j,i): jЄS}) for all subsets S of N\{i} A set S that maximizes the right-hand side of above expression is called a best response for agent i to the network g In a Nash network all agents are linked to their best responses If cik > ∑j≠ivijfor all agents k≠i, then the only best response for agent i is the empty set Ф

24. Homogeneous Costs For each agent i all links are equally expensive: cij = cifor allj

25. Homogeneous Costs For each agent i all links are equally expensive: cij = cifor allj Observation for Homogeneous Costs If link (j,k) exists in g, then for agent i ≠ j,k, linking with k is at least as good as linking with j g ●j i● k●

26. Homogeneous Costs For each agent i all links are equally expensive: cij = cifor allj Observation for Homogeneous Costs If link (j,k) exists in g, then for agent i ≠ j,k, linking with k is at least as good as linking with j g ●j i● k●

27. Homogeneous Costs For each agent i all links are equally expensive: cij = cifor allj Observation for Homogeneous Costs If link (j,k) exists in g, then for agent i ≠ j,k, linking with k is at least as good as linking with j “Downstream Efficiency” g ●j i● k●

28. Lemma For any network formation game (N,v,c) with homogeneous costs and with ci ≤ ∑j≠ivijfor all agents i, all cycle networks are Nash networks 6● ●5 1● ●4 2● 3●

29. Lemma For any network formation game (N,v,c) with homogeneous costs and with ci ≤ ∑j≠ivijfor all agents i, all cycle networks are Nash networks 6● ●5 1● ●4 2● 3●

30. Lemma For any network formation game (N,v,c) with homogeneous costs and with ci ≤ ∑j≠ivijfor all agents i, all cycle networks are Nash networks When removing (2,1) agent 1 looses profits from agents 2, 3, 4, 5, 6 6● ●5 1● ●4 2● 3●

31. Lemma For any network formation game (N,v,c) with homogeneous costs and with ci ≤ ∑j≠ivijfor all agents i, all cycle networks are Nash networks When adding (4,1) agent 1 pays an additional cost of c14 6● ●5 1● ●4 2● 3●

32. Lemma For any network formation game (N,v,c) with homogeneous costs and with ci ≤ ∑j≠ivijfor all agents i, all cycle networks are Nash networks When replacing (2,1) by (4,1) agent 1 looses profits from agents 2 and 3 6● ●5 1● ●4 2● 3●

33. Theorem For any network formation game (N,v,c) with homogeneous costs, a Nash network exists

34. Theorem For any network formation game (N,v,c) with homogeneous costs, a Nash network exists Proof by induction to the number of agents:

35. Theorem For any network formation game (N,v,c) with homogeneous costs, a Nash network exists Proof by induction to the number of agents: If n=1, then the trivial network is a Nash network

36. Theorem For any network formation game (N,v,c) with homogeneous costs, a Nash network exists Proof by induction to the number of agents: If n=1, then the trivial network is a Nash network Induction hypothesis: Nash networks exist for all network games with less than n agents.

37. Theorem For any network formation game (N,v,c) with homogeneous costs, a Nash network exists Proof by induction to the number of agents: If n=1, then the trivial network is a Nash network Induction hypothesis: Nash networks exist for all network games with less than n agents. Suppose that (N,v,c) is a network game with n agents for which NO Nash network exists.

38. Recall the Lemma: For any network formation game (N,v,c) with homogeneous costs and with ci ≤ ∑j≠ivijfor all agents i, all cycle networks are Nash networks 6● ●5 1● ●4 2● 3●

39. Proof continued: Hence there is at least one agent iwith ci > ∑j≠ivij

40. Proof continued: Hence there is at least one agent iwith ci > ∑j≠ivij W.l.o.g. this agent is agent n

41. Proof continued: Hence there is at least one agent iwith ci > ∑j≠ivij W.l.o.g. this agent is agent n Consider (N’,v’,c’)with N’=N\{n} and with v and c restricted to agents in N’

42. Proof continued: Hence there is at least one agent iwith ci > ∑j≠ivij W.l.o.g. this agent is agent n Consider (N’,v’,c’)with N’=N\{n} and with v and c restricted to agents in N’ Let g’ be a Nash network in (N’,v’,c’)(induction hypothesis)

43. Proof continued: Hence there is at least one agent iwith ci > ∑j≠ivij W.l.o.g. this agent is agent n Consider (N’,v’,c’)with N’=N\{n} and with v and c restricted to agents in N’ Let g’ be a Nash network in (N’,v’,c’)(induction hypothesis) Then by assumption g’ is no Nash network in (N,v,c)

44. Proof continued: Hence there is at least one agent iwith ci > ∑j≠ivij W.l.o.g. this agent is agent n Consider (N’,v’,c’)with N’=N\{n} and with v and c restricted to agents in N’ Let g’ be a Nash network in (N’,v’,c’)(induction hypothesis) Then by assumption g’ is no Nash network in (N,v,c) Therefore there is an agent i for whom the links in g are no best response

45. Proof continued: Hence there is at least one agent iwith ci > ∑j≠ivij W.l.o.g. this agent is agent n Consider (N’,v’,c’)with N’=N\{n} and with v and c restricted to agents in N’ Let g’ be a Nash network in (N’,v’,c’)(induction hypothesis) Then by assumption g’ is no Nash network in (N,v,c) Therefore there is an agent i for whom the links in g are no best response This agent i can not be agent n so w.l.o.g. this agent is agent 1

46. Proof continued: Hence there is at least one agent iwith ci > ∑j≠ivij W.l.o.g. this agent is agent n Consider (N’,v’,c’)with N’=N\{n} and with v and c restricted to agents in N’ Let g’ be a Nash network in (N’,v’,c’)(induction hypothesis) Then by assumption g’ is no Nash network in (N,v,c) Therefore there is an agent i for whom the links in g are no best response This agent i can not be agent n so w.l.o.g. this agent is agent 1 and he has a best response T with nЄT

47. Proof continued: Hence there is at least one agent iwith ci > ∑j≠ivij W.l.o.g. this agent is agent n Consider (N’,v’,c’)with N’=N\{n} and with v and c restricted to agents in N’ Let g’ be a Nash network in (N’,v’,c’)(induction hypothesis) Then by assumption g’ is no Nash network in (N,v,c) Therefore there is an agent i for whom the links in g are no best response This agent i can not be agent n so w.l.o.g. this agent is agent 1 and he has a best response T with nЄT and therefore c1 ≤ vin

48. Proof continued: Hence there is at least one agent iwith ci > ∑j≠ivij W.l.o.g. this agent is agent n Consider (N’,v’,c’)with N’=N\{n} and with v and c restricted to agents in N’ Let g’ be a Nash network in (N’,v’,c’)(induction hypothesis) Then by assumption g’ is no Nash network in (N,v,c) Therefore there is an agent i for whom the links in g are no best response This agent i can not be agent n W.l.o.g. this agent is agent 1 and he has a best response T with nЄT and therefore c1 ≤ vin ●n 1●

49. Proof continued: Now recall that for any other agent i linking to agent 1 would be at least as good as linking to agent n vij for j≠1 Define vij*=vi1 + vinfor i≠1, j=1 v11 + v1nfor i=1 ●n i● 1●

50. Proof continued: Now recall that for any other agent i linking to agent 1 would be at least as good as linking to agent n vij for j≠1 Define vij*=vi1 + vinfor i≠1, j=1 v11 + v1nfor i=1 Now π*i(g) = πi(g + (n,1)) for any network g on N’ and for any agent i in N’ ●n i● 1●