Strategic Network Formation and Group Formation Elliot Anshelevich Rensselaer Polytechnic Institute (RPI) Centralized Control A majority of network research has made the centralized control assumption: Everything acts according to a centrally defined and specified algorithm This assumption does not make sense in many cases. Self-Interested Agents • Internet is not centrally controlled • Many other settings have self-interested agents • To understand these, cannot assume centralized control • Algorithmic Game Theory studies such networks Agents in Network Design • Traditional network design problems are centrally controlled • What if network is instead built by many self-interested agents? • Properties of resulting network may be very different from the globally optimum one s Goal s • Compare networks created by self-interested agents with the optimal network – optimal = cheapest – networks created by self-interested agents = Nash equilibria OPT • Can realize any Nash equilibrium by finding it, and suggesting it to players – Requires central coordination – Does not require central control NE The Price of Stability Price of Anarchy = [Koutsoupias, Papadimitriou] cost(worst NE) s cost(OPT) 1 Price of Stability = k cost(best NE) cost(OPT) Can think of latter as a network designer proposing a solution. t1…tk Single-Source Connection Game [A, Dasgupta, Tardos, Wexler 2003] Given: G = (V,E), k terminal nodes, costs ce for all e E s Each player wants to build a network in which his node is connected to s. Each player selects a path, pays for some portion of edges in path (depends on cost sharing scheme) Goal: minimize payments, while fulfilling connectivity requirements Other Connectivity Requirements Survivable: connect to s with two disjoint paths [A, Caskurlu 2009] Sets of nodes: agent i wants to connect set Ti [A, Dasgupta, Tardos, Wexler 2003] Group formation Group Network Formation Games Terminal Backup: Each terminal wants to connect to k other terminals. Group Network Formation Games Terminal Backup: Each terminal wants to connect to k other terminals. “Group Steiner Tree”: Each terminal wants to connect to at least one terminal from each color. Other Connectivity Requirements Survivable: connect to s with two disjoint paths [A, Caskurlu 2009] Sets of nodes: agent i wants to connect set Ti [A, Dasgupta, Tardos, Wexler 2003] Group formation: every agent wants to connect to a group that provides enough resources satisfactory group specified by a monotone set function [A, Caskurlu 2009] Centralized Optimum Single-source Connection Game: Steiner Tree. Sets of nodes: Steiner Forest. Survivable: Generalized Steiner Forest. Terminal Backup: Cheapest network where each terminal connected to at least k other terminals. “Group Steiner Tree”: Cheapest where every component is a Group Steiner Tree. Corresponds to constrained forest problems, has 2-approx. Connection Games s Given: G = (V,E), k players, costs ce for all e E Each player wants to build a network where his connectivity requirements are satisfied. Each player selects subgraph, pays for some portion of edges in it (depends on cost sharing scheme) NE Goal: minimize payments, while fulfilling connectivity requirements Sharing Edge Costs How should multiple players on a single edge split costs? One approach: no restrictions... ...any division of cost agreed upon by players is OK. [ADTW 2003, HK 2005, EFM 2007, H 2009, AC 2009] Another approach: try to ensure some sort of fairness. [ADKTWR 2004, CCLNO 2006, HR 2006, FKLOS 2006] Connection Games with Fair Sharing Given: G = (V,E), k players, costs ce for all e E s Each player selects subnetwork where his connectivity requirements are satisfied. Players using e pay for it evenly: ci(P) = Σ ce/ke eєP ( ke = # players using e ) Goal: minimize payments, while fulfilling connectivity requirements Fair Sharing Fair sharing: The cost of each edge e shared equally by the users of e Advantages: • Fair way of sharing the cost • Nash equilibrium exists • Price of Stability is at most log(# players) is Price of Stability with Fairness Price of Anarchy is large Price of Stability is at most log(# players) s Proof: This is a Potential Game, so Nash equilibrium exists Best Response converges Can use this to show existence of good equilibrium 1 k t1…tk Fair Sharing Fair sharing: The cost of each edge e shared equally by the users of e Advantages: • Fair way of sharing the cost • Nash equilibrium exists • Price of Stability is at most log(# players) Disadvantages: • Player payments are constrained, need to enforce fairness • Price of stability can be at least log(# players) is Example: Self-Interested Behavior t 1 1 1+ 1 2 1 Demands: 1-t, 2-t, 3-t 3 2 0 0 3 0 Example: Self-Interested Behavior t 1 1 1+ 1 2 1 Minimum Cost Solution (of cost 1+) 3 2 0 0 3 0 Example: Self-Interested Behavior Each player chooses a path P. Cost to player i is: t 1 1 1+ 1 2 1 2 0 0 cost(i) = cost(P) # using P 3 3 0 (Everyone shares cost equally) Example: Self-Interested Behavior t 1 1 1+ 1 2 1 3 2 0 0 3 0 Player 3 pays (1+ε)/3, could pay 1/3 Example: Self-Interested Behavior t 1 1 1+ 1 2 1 3 2 0 0 3 0 so player 3 would deviate Example: Self-Interested Behavior t 1 1 1+ 1 2 1 3 2 0 0 3 0 now player 2 pays (1+ε)/2, could pay 1/2 Example: Self-Interested Behavior t 1 1 1+ 1 2 1 3 2 0 0 3 0 so player 2 deviates also Example: Self-Interested Behavior Player 1 deviates as well, giving a solution with cost 1.833. This solution is stable/ this solution is a Nash Equilibrium. t 1 1 1+ 1 2 1 3 2 0 0 3 0 It differs from the optimal solution by a factor of 1+ 12 + 13 Hk = Θ(log k)! Sharing Edge Costs How should multiple players on a single edge split costs? One approach: no restrictions... ...any division of cost agreed upon by players is OK. [ADTW 2003, HK 2005, EFM 2007, H 2009, AC 2009] Another approach: try to ensure some sort of fairness. [ADKTWR 2004, CCLNO 2006, HR 2006, FKLOS 2006] Example: Unrestricted Sharing Fair Sharing: differs from the optimal solution by a factor of Hk = Θ(log k) t 1 1 1+ 1 2 1 3 2 0 0 3 0 Unrestricted Sharing: OPT is a stable solution Contrast of Sharing Schemes Unrestricted Sharing Fair Sharing NE don’t always exist P.o.S. = O(k) NE always exist P.o.S. = O(log(k)) (P.o.S. = Price of Stability) Contrast of Sharing Schemes Unrestricted Sharing Fair Sharing NE don’t always exist P.o.S. = O(k) P.o.S. = 1 for many games NE always exist P.o.S. = O(log(k)) P.o.S. = (log(k)) for almost all games (P.o.S. = Price of Stability) Contrast of Sharing Schemes Unrestricted Sharing Fair Sharing NE don’t always exist P.o.S. = O(k) P.o.S. = 1 for many games OPT is an approx. NE NE always exist P.o.S. = O(log(k)) P.o.S. = (log(k)) for almost all games OPT may be far from NE (P.o.S. = Price of Stability) Unrestricted Sharing Model • Player i picks payments for each edge e. (strategy = vector of payments) • Edge e is bought if total payments for it ≥ ce. • Any player can use bought edges. What is a NE in this model? Unrestricted Sharing Model • Player i picks payments for each edge e. (strategy = vector of payments) • Edge e is bought if total payments for it ≥ ce. • Any player can use bought edges. What is a NE in this model? Payments so that no players want to change them Unrestricted Sharing Model • Player i picks payments for each edge e. (strategy = vector of payments) • Edge e is bought if total payments for it ≥ ce. • Any player can use bought edges. What is a NE in this model? Payments so that no players want to change them Connection Games with Unrestricted Sharing Given: G = (V,E), k players, costs ce for all e E s Strategy: a vector of payments Players choose how much to pay, buy edges together Cost(v) = if v does not satisfy connectivity requirements Payments of v otherwise Goal: minimize payments, while fulfilling connectivity requirements Connectivity Requirements Single-source: connect to s Survivable: connect to s with two disjoint paths Sets of nodes: agent i wants to connect set Ti Group formation: every agent wants to connect to a group that provides enough resources satisfactory group specified by a monotone set function Some Results Single-source: connect to s Survivable: connect to s with two disjoint paths If k=n Sets of nodes: agent i wants to connect set Ti Group formation: every agent wants to connect to a group that provides enough resources satisfactory group specified by a monotone set function If k=n OPT is a Nash Equilibrium (Price of Stability=1) Some Results Single-source: connect to s =1 Survivable: connect to s with two disjoint paths =2 Sets of nodes: agent i wants to connect set Ti =3 Group formation: every agent wants to connect to a group that provides enough resources satisfactory group specified by a monotone set function =2 OPT is a -approximate Nash Equilibrium (no one can gain more than factor by switching) Some Results Single-source: connect to s =1 Survivable: connect to s with two disjoint paths =2 Sets of nodes: agent i wants to connect set Ti =3 Group formation: every agent wants to connect to a group that provides enough resources satisfactory group specified by a monotone set function =2 If we pay for 1-1/ fraction of OPT, then the players will pay for the rest Some Results Single-source: connect to s Survivable: connect to s with two disjoint paths Sets of nodes: agent i wants to connect set Ti Group formation: every agent wants to connect to a group that provides enough resources satisfactory group specified by a monotone set function Can compute cheap approximate equilibria in poly-time Contrast of Sharing Schemes Unrestricted Sharing Fair Sharing NE don’t always exist P.o.S. = O(k) P.o.S. = 1 for many games OPT is an approx. NE NE always exist P.o.S. = O(log(k)) P.o.S. = (log(k)) for almost all games OPT may be far from NE (P.o.S. = Price of Stability) Contrast of Sharing Schemes Unrestricted Sharing Fair Sharing NE don’t always exist P.o.S. = O(k) P.o.S. = 1 for many games OPT is an approx. NE NE always exist P.o.S. = O(log(k)) P.o.S. = (log(k)) for almost all games OPT may be far from NE (P.o.S. = Price of Stability) Contrast of Sharing Schemes Unrestricted Sharing Fair Sharing NE don’t always exist P.o.S. = O(k) P.o.S. = 1 for many games OPT is an approx. NE NE always exist P.o.S. = O(log(k)) P.o.S. = (log(k)) for almost all games OPT may be far from NE If we really care about efficiency: Allow the players more freedom! Example: Unrestricted Sharing Fair Sharing: differs from the optimal solution by a factor of Hk log k t 1 1 1+ 1 2 1 3 2 0 0 3 0 Unrestricted Sharing: OPT is a stable solution Every player gives what they can afford General Techniques To prove that OPT is an exact/approximate equilibrium: Construct a payment scheme Pay in order: laminar system of witness sets If cannot pay, form deviations to create cheaper solution Te T' u e pi i ( pi ) Te Network Destruction Games • Each player wants to protect itself from untrusted nodes • Have cut requirements: must be disconnected from set Ti • Cutting edges costs money • Can show similar results for: Multiway Cut, Multicut, etc. Thank you.