Resource Allocation Algorithms

Transcription

1 Resource Allocation Algorithms Haris Aziz 1, 2 1 School of Computer Science and Engineering, UNSW Australia 2 Data61, CSIRO April, 2018 H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

2 Outline 1 Allocation setting 2 Fair Allocation Under Additive Utilities 3 Allocation under Endowments H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

3 Outline 1 Allocation setting 2 Fair Allocation Under Additive Utilities 3 Allocation under Endowments H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

4 Allocation Setting Basic Allocation Setting Agents N = {1,..., n} Items O = {o 1,..., o m} Preferences (of agents) = { 1,..., n}; preferences can be encoded by utility function u = (u 1,..., u n) over bundles of items. An allocation X = (X(1),..., X(n)) assigns X(i) O to agent i. We will assume that X(i) X(j) = for all i, j N such that i j. We will focus on allocations that allocate all the items: i N Xi = O. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

5 Some notation A i B (agent i prefers A at least as much as B) A i B A i B and B i A (agent i strictly prefers A over B) A i B A i B and B i A (agent i is indifferent between A and B). H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

6 Allocation setting: Additive Utilities We assume additive utilities: u i : O R + specifies the utility function of each agent i. u i (O ) = o O u i(o) for any O O. Example o 1 o 2 o 3 o u 1 (o 1 ) = 6; u 1 (o 2 ) = 3; u 1 (o 3 ) = 2; u 1 (o 4 ) = 1. u 1 ({o 1, o 2 }) > u 1 ({o 2, o 3 }). {o 1, o 2 } 1 {o 2, o 3 }. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

7 Outline 1 Allocation setting 2 Fair Allocation Under Additive Utilities 3 Allocation under Endowments H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

8 Envy-freeness An allocation X satisfies envy-freeness if for all i, j N X(i) i X(j) u i (X(i)) u i (X(j)) Was formally introduced by Foley (1967). Example (Not envy-free) X(1) = {o 1, o 2, o 3 }, X(2) = {o 4 }. o 1 o 2 o 3 o H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

9 Proportional An allocation X satisfies proportionality if for all i N Example (Not proportional) X(1) = {o 1, o 2, o 3 }, X(2) = {o 4 }. u i (X(i)) u i(o) n o 1 o 2 o 3 o H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

10 Envy-freeness implies proportionality Fact If an allocation is complete and utilities are additive, envy-freeness implies proportionality. Assume that an allocation X is envy-free. Then for each i N, u i (X(i)) u i (X(j)) for all j N. Thus, n u i (X(i)) j N u i (X(j)) = u i (O). Hence u i (X(i)) u i (O)/n. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

11 Non-existence of envy-free or proportional allocation Example o 1 o H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

12 Allocation of indivisible items Theorem (Demko and Hill [1988]) For additive utilities, checking whether there exists an envy-free or proportional allocation is NP-complete. Proof. We present a reduction from the following NP-complete problem. IntegerPartition Input: A set of integers S = {w 1,..., w m } such that w S w = 2W. Question: Does there exist a partiton (S, S ) of S such that w S w = w S w = W? Consider the setting in which two agents have identical utilities over the m items with the utility for the j-th item being w j and the total utility of each agents over the items being 2W. Then, there exists a proportional allocation iff there is an integer partition of the integers corresponding to the weights so that each partition has total weight W. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

13 EF1 Fairness Definition (EF1 Fairness) Given an instance I = (N, O, u), an allocation X satisfies EF1 (envy-freeness up to 1 item) if for each i, j N, there exists some item o X(j) such that Example (Satisfies EF1 Fairness) X(1) = {o 1, o 2, o 3 }, X(2) = {o 4 }. X(i) i X(j) \ {o}. o 1 o 2 o 3 o H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

14 Algorithm for EF1 fairness (Lipton et al. (2004) ) Algorithm by Lipton, Markakis, Mossel, and Saberi [2004] Envy graph (an agent points to another agent if she envies her). Suppose the graph is for a partial allocation that is EF1 fair. Agent 5 has no incoming arc so if she gets a new item, the allocation is still EF1 fair. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

15 Algorithm for EF1 fairness (Lipton et al. (2004) ) A new item is given to agent 5 who has no incoming arc. This may make some other agent envious (in this case agent 3 is now envious of agent 5). H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

16 Algorithm for EF1 fairness (Lipton et al. (2004) ) We enable an exchange of allocations along the cycle which removes the cycle. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

17 Algorithm for EF1 fairness (Lipton et al. (2004) ) We enable an exchange of allocations along the cycle which removes the cycle. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

18 Algorithm for EF1 fairness (Lipton et al. (2004) ) Input : n agents, m items, and valuations u i (o j ) for each i [n] and j O. Output: EF1 allocation X 1: Initialize allocation X = (X(1), X(2),..., X(n)) with X(i) = for all i [n]. 2: for j = 1 to m do 3: Construct an envy-graph G(X) = (N, E) where (i, j) E if i is envious of j s allocation wrt allocation X. 4: Pick a vertex i that has no incoming edges in G(X) 5: Update X(i) X(i) {o j }. 6: while the G(X) contains a cycle do 7: Implement an exchange in which if i points to j in the cycle, then i gets j s allocation. 8: end while 9: end for 10: Return X. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

19 Fairness Overview EF implies proportionality. EF implies EF1 fairness. EF and Proportional fair allocations may not exist and are computationally hard to compute even if they exist. An EF1 allocation always exists and can be computed in polynomial time. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

20 Outline 1 Allocation setting 2 Fair Allocation Under Additive Utilities 3 Allocation under Endowments H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

21 Housing market: model with endowments (N, O,, ω) N = O ω(i) = {o} iff o is owned by i N. Agents have strict preferences over items Each agent owns and is allocated one item. Example Housing market (N, O, ω, ) such that N = {1,..., 5}, O = {o 1,..., o 5 }, ω(i) = {o i } for all i {1,..., 5} and preferences are defined as follows: agent preferences o 2 o 3 o 4 o 1 o 2 o 1 o 2 o 3 o 5 o 4 o 4 o 5 H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

22 Individual rationality An allocation X is individually rational if no agent minds participating in the allocation procedure: i N : X(i) i ω(i) If an agent does not have any endowment, her allocation is individually rational if her allocation is acceptable (at least as preferred as the empty allocation). An agent can express an allocation or an item as unacceptable by simply not listing it in the preference list. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

23 Allocation with endowments: Core An allocation X is core stable if there exists no S N such that there exists an allocation Y of the items in i S ω(i) to the agents in S such that i S : Y (i) i X(i) Fact A core stable allocation is individually rational. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

24 Housing Markets: Gale s Top Trading Cycles (TTC) Algorithm Each item points to its owner. Each agent points to her most preferred item in the graph. Find a cycle, allocate to each agent in the cycle the item she was pointing to. Remove the agents and items in the cycle. Adjust the graph so the agents in the graph point to their most preferred item in the graph. Repeat until the graph is empty. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

25 Housing Markets: Gale s Top Trading Cycles (TTC) Algorithm Each item points to its owner. Each agent points to her most preferred item in the graph. Find a cycle, allocate to each agent in the cycle the item she was pointing to. Remove the agents and items in the cycle. Adjust the graph so the agents in the graph point to their most preferred item in the graph. Repeat until the graph is empty. agents 1 2 item owned o 1 o 2 1 agents 1 2 preferences o 2 o 1 o 2 o 1 o 1 o 2 2 H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

26 Housing Markets: Gale s Top Trading Cycles (TTC) Algorithm Each item points to its owner. Each agent points to her most preferred item in the graph. Find a cycle, allocate to each agent in the cycle the item she was pointing to. Remove the agents and items in the cycle. Adjust the graph so the agents in the graph point to their most preferred item in the graph. Repeat until the graph is empty. agents 1 2 item owned o 1 o 2 1 agents 1 2 preferences o 2 o 1 o 2 o 1 o 1 o 2 2 H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

27 Housing Market Example Example Housing market M = (N, O, ω, ) such that N = {1,..., 5}, O = {o 1,..., o 5 }, ω(i) = {o i } for all i {1,..., 5} and preferences are defined as follows: agent preferences o 2 o 3 o 4 o 1 o 2 o 1 o 2 o 3 o 5 o 4 o 4 o 5 H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

28 Housing Markets: Gale s Top Trading Cycles (TTC) Algorithm agent o 3 2 o 2 1 preferences o 2 o 3 o 4 o 1 o 2 3 o o 1 o 2 o 3 o 5 o 1 4 o 4 o 5 o 4 4 o 5 5 H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

29 Example: TTC agent o 3 2 o 2 1 preferences o 2 o 3 o 4 o 1 o 2 3 o o 1 o 2 o 3 o 5 o 1 4 o 4 o 5 o 4 4 o 5 5 H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

30 Housing Markets: Gale s Top Trading Cycles (TTC) Algorithm agent o 3 2 o 2 1 preferences o 2 o 3 o 4 o 1 o 2 3 o o 1 o 2 o 3 o 5 o 1 4 o 4 o 5 o 4 4 o 5 5 H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

31 Housing Markets: Gale s Top Trading Cycles (TTC) Algorithm agent o 3 2 o 2 1 preferences o 2 o 3 o 4 o 1 o 2 3 o o 1 o 2 o 3 o 5 o 1 4 o 4 o 5 o 4 4 o 5 5 H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

32 TTC (Top Trading Cycles) Theorem (Shapley and Scarf [1974]) For housing markets (with strict preferences), TTC finds a core stable allocation. H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

33 Survey and Further Reading Most relevant resource: book chapter by Bouveret, Chevaleyre, and Maudet [2016] in the Handbook of Computational Social Choice. Brandt, Conitzer, Endriss, Lang, and Procaccia [2016] especially chapters Brams and Taylor [1996] Robertson and Webb [1998] Moulin [2003] Endriss [2010] Roth and Sotomayor [1990] Gusfield and Irving [1989] Manlove [2013] Chalkiadakis, Elkind, and Wooldridge [2011] H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

34 Contact H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

35 References I S. Bouveret, Y. Chevaleyre, and N. Maudet. Fair allocation of indivisible goods. In F. Brandt, V. Conitzer, U. Endriss, J. Lang, and A. D. Procaccia, editors, Handbook of Computational Social Choice, chapter 12. Cambridge University Press, S. J. Brams and A. D. Taylor. Fair Division: From Cake-Cutting to Dispute Resolution. Cambridge University Press, F. Brandt, V. Conitzer, U. Endriss, J. Lang, and A. Procaccia, editors. Handbook of Computational Social Choice. Cambridge University Press, G. Chalkiadakis, E. Elkind, and M. Wooldridge. Computational Aspects of Cooperative Game Theory, volume 5 of Synthesis Lectures on Artificial Intelligence and Machine Learning. Morgan & Claypool, S. Demko and T. P. Hill. Equitable distribution of indivisible objects. Mathematical Social Sciences, 16: , U. Endriss. Lecture notes on fair division D. Gusfield and R. W. Irving. The stable marriage problem: Structure and algorithms. MIT Press, Cambridge, MA, USA, H. Aziz (UNSW) Resource Allocation Algorithms April, / 33

36 References II R. J. Lipton, E. Markakis, E. Mossel, and A. Saberi. On approximately fair allocations of indivisible goods. In Proceedings of the 5th ACM Conference on Electronic Commerce (ACM-EC), pages ACM Press, D. Manlove. Algorithmics of Matching Under Preferences. World Scientific Publishing Company, H. Moulin. Fair Division and Collective Welfare. The MIT Press, J. M. Robertson and W. A. Webb. Cake Cutting Algorithms: Be Fair If You Can. A. K. Peters, A. E. Roth and M. A. O. Sotomayor. Two-Sided Matching: A Study in Game Theoretic Modelling and Analysis. Cambridge University Press, L. S. Shapley and H. Scarf. On cores and indivisibility. Journal of Mathematical Economics, 1(1):23 37, H. Aziz (UNSW) Resource Allocation Algorithms April, / 33