Extended Abstract AAMAS 2022, May 9–13, 2022, Online

Concise Representations and Complexity of Combinatorial

Assignment Problems
Extended Abstract
Fredrik Präntare George Osipov Leif Eriksson
Linköping University Linköping University Linköping University
Linköping, Sweden Linköping, Sweden Linköping, Sweden
[email protected] [email protected] [email protected]

ABSTRACT Examples include multi-agent scenarios where the goal is to

We consider the computational problem of partitioning items into simultaneously form agent teams and assign them to tasks.
bundles among alternatives to maximize social welfare. Unfortu- (2) Egalitarian combinatorial assignment (ECA), in which the
nately, many important classes of this problem are computationally value of the bundle that is worst off is maximized; i.e., we
hard, including well-known instances in the multi-agent systems lit- want to keep the weakest link as strong as possible. An exam-
erature. In this paper we analyze novel concise representations and ple of this is equitable distribution of indivisible resources.
restrictions that admit polynomial-time algorithms for many such To this end, our main contributions can be summarized as follows:
combinatorial assignment problems, and prove several complexity
results for them. We provide efficient approximation algorithms • We develop a novel hypergraph characterization of UCA,
and non-trivial exponential-time algorithms for the hard cases. enabling us to prove several new state-of-the-art results.
• We consider practical restrictions that make combinatorial
KEYWORDS assignment problems tractable and/or approximable in poly-
nomial time. This work includes e.g., bounded combinatorial
Combinatorial Assignment; Complexity; Resource Allocation; Coali- assignment, in which bundles of size larger than some con-
tion Structure Generation; Combinatorial Auctions stant are given value zero.
ACM Reference Format: • We explore combinatorial assignment generalizations with
Fredrik Präntare, George Osipov, and Leif Eriksson. 2022. Concise Represen- externalities (i.e., cross-bundle value dependencies) that we
tations and Complexity of Combinatorial Assignment Problems: Extended analyze and develop exponential-time algorithms for.
Abstract. In Proc. of the 21st International Conference on Autonomous Agents
• We investigate the balanced (also called mixed) egalitarian
and Multiagent Systems (AAMAS 2022), Online, May 9–13, 2022, IFAAMAS,
and utilitarian combinatorial assignment problem.
3 pages.

1 INTRODUCTION 2 RELATED WORK

Efficiently distributing resources in a sustainable and effective way We first turn to the winner determination problem (WDP) for com-
is one of the most important and challenging problems in society. binatorial auctions—a canonical combinatorial assignment prob-
Different combinatorial assignment problems, in which the goal is to lem in which the goal is to maximize an auctioneer’s profit. For
divide a number of indivisible items among alternatives to maximize the WDP, [24] provided several complexity bounds for various
welfare/value, appear in a wide range of settings. Such settings application-focused restrictions. [14] analyzed how the WDP’s
include operations research, economics and artificial intelligence, complexity is affected by using bidding languages, and [11] stud-
with real-world applications in combinatorial auctions [27], multi- ied restrictions to hypergraphs describing the bids. [9] studied the
target tracking and sensor fusion [8], course assignment [4, 19], WDP’s complexity when the goal is instead to find an equilibrium.
resource allocation [15], and team/coalition formation [16, 21]. For coalition structure generation (CSG) [26]—in which agents are
Many important combinatorial assignment problems are how- partitioned into subsets called coalitions—[1, 2] considered different
ever computationally hard, hard to approximate in polynomial time, compactly representable problems. [31] analyzed CSG over graphs,
and/or require an input size that is exponential in one of the pa- and [17] investigated rule-based representations.
rameters. Important examples of such problems include coalition Related to our work on cross-bundle dependencies is CSG with
structure generation [26], combinatorial auction winner determina- externalities, which adds cross-coalitional effects to the coalitional
tion [25], and generalized assignment [5]. values [23]. For this problem, only a few concise representations
In light of this, we investigate the computational aspects of two have so far been considered, such as the tree-based one by [32].
hard general classes of combinatorial assignment problems: Our work on externalities can also be applied to such problems.
(1) Utilitarian combinatorial assignment (UCA), in which the Another important combinatorial assignment problem is the gen-
sum of all the bundle-to-alternative assignments is maxi- eralized assignment problem (GAP) [10], which [5] surveyed. In the
mized with the goal to maximize a system’s total output. GAP, all items are given weights unique to the different alternatives,
and the alternatives have capacities that denote the maximum total
Proc. of the 21st International Conference on Autonomous Agents and Multiagent Systems weight of items that can be assigned to them. The corresponding
(AAMAS 2022), P. Faliszewski, V. Mascardi, C. Pelachaud, M.E. Taylor (eds.), May 9–13,
2022, Online. © 2022 International Foundation for Autonomous Agents and Multiagent value function is additive, so there are no synergies, making the
Systems (www.ifaamas.org). All rights reserved. problem approximable [29] and more similar to knapsack problems.

1714
Extended Abstract AAMAS 2022, May 9–13, 2022, Online

Table 1: Results for the 𝑘-bounded case. Here 𝑙 = max(𝑛, 𝑚). 4 RESULTS AND CONCLUSIONS
We analyzed restrictions and concise representations for several
𝑘=1 𝑘≥2 𝑚≥2 𝑚 and 𝑘 fixed combinatorial assignment problems—the problems in which the
UCA O (𝑙 3𝑚) NP-hard NP-hard (3𝑛)𝑘𝑚 O (𝑚) goal is to partition 𝑛 items into bundles among 𝑚 alternatives to
p
maximize social welfare. We considered both utilitarian (i.e., maxi-
ECA O 𝑚 𝑙 5 log(𝑙) NP-hard NP-hard (3𝑛)𝑘𝑚 O (𝑘𝑚 2 )
mize the total output) and egalitarian (i.e., maximize the worst-case
allocation) welfare functions. For various concise representations,
generalizations and hypergraph characterizations of them, we pro-
Similar to our 𝑘-bounded combinatorial assignment—in which
vided novel algorithms and new complexity results.
bundles of size larger than 𝑘 are given value zero—are 𝑘-additive
In our hypergraph characterization, the value function is repre-
domains [7]. Specifically, 𝑘-additive problems are the UCA instances
sented using a hypergraph called the synergy hypergraph, and we
that can be represented by hypergraphs with hyperedges of size
denoted the corresponding problem UCA★. For UCA★, the value of
at most 𝑘 + 1. Moreover, [28] proved several complexity results for
a bundle equals the sum of all of the (non-negative) synergies for
quadratic value functions, which is equivalent to 2-bounded UCA.
every subset of the items in the bundle. We found that the case with
For ECA and mixed welfare, [30] outlines corresponding egali-
two alternatives admits a polynomial-time algorithm by comput-
tarian and balanced versions of CSG, and [6] presents maximizing
ing a cut of minimum cost in the hypergraph using an algorithm
egalitarian welfare as a way to offer a level of fairness in a system.
by [13]. However, more alternatives lead to intractability.
Finally, [18] studied a course assignment problem in which the
goal is to find an equilibrium, and [3] provided complexity results Theorem 4.1. UCA★ with 𝑚 = 2 is solvable in polynomial time,
for a related fair division resource allocation problem. while UCA★ with 𝑚 ≥ 3 is NP-hard.
Using this result, we were able to construct improved (faster than
3 COMBINATORIAL ASSIGNMENT DP’s O ∗ (3𝑛 )) exp-time algorithms for some of the hard cases.
In this paper we consider combinatorial assignment—i.e., the class
of problems in which a set 𝑁 = {1, . . . , 𝑛} of indivisible items (e.g., Theorem 4.2. UCA★ with 𝑚 = 3, 𝑚 = 4 and 𝑚 = 5 is solvable in
goods/agents) have to be distributed in bundles (i.e., partitioned) O ∗ (1.89𝑛 ), O ∗ (2𝑛 ) and O ∗ (2.89𝑛 ) time, respectively.
among a set 𝑀 = {1, . . . , 𝑚} of alternatives (e.g., buyers/jobs) to Under the aforementioned 𝑘-bounded restriction, we found that
maximize social welfare. The resulting ordered size-𝑚 partition of the problem remains NP-hard to approximate to some constant
such a problem is called a combinatorial assignment. in polynomial time. On the positive side, 𝑘-bounded non-negative
Definition 3.1. The tuple (𝐵 1, . . . , 𝐵𝑚 ) is a combinatorial as- UCA can be approximated within 𝑘+2 2 of the optimum in polynomial
Ð
signment over 𝑁 if 𝐵𝑖 ∩ 𝐵 𝑗 = ∅ for all 𝑖 ≠ 𝑗, and 𝑚 𝑖=1 𝐵𝑖 = 𝑁 .
time. Table 1 summarizes our other results for this representation.
Finally, we explored complexity bounds for generalizations with
We use Π 𝑁 to denote the set of all combinatorial assignments balanced utilitarian and egalitarian welfare functions (BCA), and
over 𝑁 , and define Π𝑘𝑁 = {𝐶 ∈ Π 𝑁 : |𝐶 | = 𝑘 } for 𝑘 ∈ {1, ..., 𝑚}. The externalities (XCA), for which a bundle’s value can depend on other
combinatorial assignment problem can now be defined as follows. bundles’ values. For BCA, we presented an exact O ∗ (4𝑛 ) time al-
The Combinatorial Assignment Problem gorithm. For a relaxation of XCA—denoted XCA★—we found that
Input: A set of 𝑛 items 𝑁 , a set of 𝑚 alternatives 𝑀, and a externalities makes the problem significantly harder, providing a
function (called the social welfare function) Φ : Π𝑚 𝑁 → R. lower bound under the exponential time hypothesis [12].
Output: A combinatorial assignment (𝐵 1, . . . , 𝐵𝑚 ) that maxi-
mizes the value (or social welfare) Φ((𝐵 1, . . . , 𝐵𝑚 )). Theorem 4.3. XCA★ cannot be solved in 2𝑜 (𝑛 log 𝑚) time unless
the exponential time hypothesis is false.
For UCA—the most central problem that we explore—the social
While, in general, combinatorial assignment problems are com-
welfare function is defined as
𝑚
putationally hard, we found that certain restrictions lead to tractabil-
Õ
Φ((𝐵 1, . . . , 𝐵𝑚 )) = 𝑣 (𝐵𝑖 , 𝑖), ity, allowing either polynomial-time exact and approximate algo-
𝑖=1
rithms, or non-trivial exponential-time algorithms. Thus, investi-
gating structural restrictions is an interesting future direction. One
where 𝑣 is a function 𝑣 : 2𝑁
× 𝑀 → R (called the value function). possibility is to study synergy hypergraphs whose associated pri-
Due to this welfare function, UCA is equivalent to simultaneous mal, dual, and incidence graphs have bounded treewitdh; even more
coalition structure generation and assignment [21] when the items general parameters like primal and dual hypertree-width are also in-
are viewed as agents. It also generalizes other notable problems teresting. A combination of parameters can also be considered, e.g.,
in a straightforward fashion, such as many of the aforementioned the cases with bounded bundle-size and bounded hypertree-width.
combinatorial assignment problems. UCA is NP-hard [20], and can
be solved with dynamic programming (DP) in O ∗ (3𝑛 ) [22] (the ACKNOWLEDGMENTS
notation O ∗ (·) hides polynomial factors).
This work was partially supported by the Wallenberg AI, Autonomous
For ECA (which can also be solved with DP), with the same
Systems and Software Program (WASP) funded by the Knut and
function 𝑣, the social welfare function is defined with
𝑚
Alice Wallenberg Foundation and by the National Graduate School
Φ((𝐵 1, . . . , 𝐵𝑚 )) = min 𝑣 (𝐵𝑖 , 𝑖). in Computer Science (CUGS), Sweden.
𝑖=1

1715
Extended Abstract AAMAS 2022, May 9–13, 2022, Online

REFERENCES Programming (CP). 623–638.

[1] H. Aziz and B. De Keijzer. 2011. Complexity of coalition structure generation. [18] Abraham Othman, Christos Papadimitriou, and Aviad Rubinstein. 2016. The
In Proceedings of the 10th International Conference on Autonomous Agents and complexity of fairness through equilibrium. ACM Transactions on Economics and
Multiagent Systems (AAMAS), Vol. 1. 191–198. Computation (TEAC) 4, 4 (2016), 1–19.
[2] Y. Bachrach, P. Kohli, V. Kolmogorov, and M. Zadimoghaddam. 2013. Optimal [19] A. Othman, T. Sandholm, and E. Budish. 2010. Finding approximate competitive
coalition structure generation in cooperative graph games. In 27th AAAI Confer- equilibria: efficient and fair course allocation. In Proceedings of the 9th Inter-
ence on Artificial Intelligence. national Conference on Autonomous Agents and Multiagent Systems (AAMAS),
[3] Sylvain Bouveret and Jérôme Lang. 2008. Efficiency and envy-freeness in fair Vol. 10. 873–880.
division of indivisible goods: Logical representation and complexity. Journal of [20] F. Präntare, H. Appelgren, and F. Heintz. 2021. Anytime Heuristic and Monte
Artificial Intelligence Research (JAIR) 32 (2008), 525–564. Carlo Methods for Large-Scale Simultaneous Coalition Structure Generation and
[4] E. Budish. 2011. The combinatorial assignment problem: Approximate compet- Assignment. In AAAI Conference on Artificial Intelligence, Vol. 35. 11317–11324.
itive equilibrium from equal incomes. Journal of Political Economy (JPE) 119 [21] F. Präntare and F. Heintz. 2020. An Anytime Algorithm for Optimal Simultaneous
(2011), 1061–1103. Coalition Structure Generation and Assignment. Journal of Autonomous Agents
[5] D. G. Cattrysse and L. N. Van Wassenhove. 1992. A survey of algorithms for and Multiagent Systems (JAAMAS) 34 (2020).
the generalized assignment problem. European Journal of Operational Research [22] F. Präntare and F. Heintz. 2020. Hybrid Dynamic Programming for Simultaneous
(EJOR) 60 (1992), 260–272. Coalition Structure Generation and Assignment. In International Conference on
[6] Yann Chevaleyre, Paul E Dunne, Ulle Endriss, Jérôme Lang, Michel Lemaître, Principles and Practice of Multi-Agent Systems (PRIMA). 19–33.
Nicolas Maudet, Julian Padget, Steve Phelps, Juan A Rodríguez-Aguilar, and Paulo [23] T. Rahwan, T. Michalak, N. Jennings, M. Wooldridge, and P. McBurney. 2009.
Sousa. 2006. Issues in Multiagent Resource Allocation. Hot Topics in European Coalition structure generation in multi-agent systems with positive and negative
Agent Research II Guest Editors: Andrea Omicini 30 (2006), 3–31. externalities. In 21th International Joint Conference on Artificial Intelligence (IJCAI).
[7] Yann Chevaleyre, Ulle Endriss, Sylvia Estivie, and Nicolas Maudet. 2008. Multia- 257–263.
gent resource allocation in k-additive domains: Preference representation and [24] M. H. Rothkopf, A. Pekeč, and R. M. Harstad. 1998. Computationally manageable
complexity. Annals of Operations Research 163, 1 (2008), 49–62. combinational auctions. Management science 44 (1998), 1131–1147.
[8] V. D. Dang, R. K. Dash, A. Rogers, and N. R. Jennings. 2006. Overlapping coalition [25] T. Sandholm. 2002. Algorithm for optimal winner determination in combinatorial
formation for efficient data fusion in multi-sensor networks. In AAAI. 635–640. auctions. Artificial Intelligence 135 (2002), 1–54.
[9] S. Dobzinski, H. Fu, and R. Kleinberg. 2014. On the complexity of computing an [26] T. Sandholm, K. Larson, M. Andersson, O. Shehory, and F. Tohmé. 1999. Coalition
equilibrium in combinatorial auctions. In Proceedings of the 26th Annual ACM- structure generation with worst case guarantees. Artificial intelligence 111 (1999),
SIAM Symposium on Discrete Algorithms. 110–122. 209–238.
[10] M. L. Fisher and R. Jaikumar. 1981. A generalized assignment heuristic for vehicle [27] T. Sandholm, S. Suri, A. Gilpin, and D. Levine. 2002. Winner Determination in
routing. Networks 11 (1981), 109–124. Combinatorial Auction Generalizations. In Proceedings of the 2nd International
[11] G. Gottlob and G. Greco. 2013. Decomposing combinatorial auctions and set Conference on Autonomous Agents and Multiagent Systems (AAMAS). 69–76.
[28] Akiyoshi Shioura and Shunya Suzuki. 2011. Optimal allocation in combinatorial
packing problems. J. ACM 60 (2013), 1–39.
auctions with quadratic utility functions. In International Conference on Theory
[12] Russell Impagliazzo and Ramamohan Paturi. 2001. On the complexity of k-SAT.
and Applications of Models of Computation (TAMC). Springer, 142–153.
J. Comput. System Sci. 62, 2 (2001), 367–375.
[29] D. B. Shmoys and É. Tardos. 1993. An approximation algorithm for the generalized
[13] E. L. Lawler. 1973. Cutsets and partitions of hypergraphs. Networks 3 (1973),
assignment problem. Mathematical programming 62 (1993), 461–474.
275–285.
[30] Oskar Skibski, Henryk Michalewski, Andrzej Nagórko, Tomasz P Michalak, An-
[14] D. Lehmann, R. Müller, and T. Sandholm. 2006. The winner determination
drew Dowell, Talal Rahwan, and Michael Wooldridge. 2016. Non-utilitarian
problem. Combinatorial Auctions (2006), 297–318.
coalition structure generation. In Proceedings of the Twenty-second European
[15] K. Leyton-Brown. 2003. Resource Allocation in Competitive Multiagent Systems.
Conference on Artificial Intelligence. 1738–1739.
Ph.D. Dissertation. Stanford University.
[31] T. Voice, M. Polukarov, and N. R. Jennings. 2012. Coalition structure generation
[16] S. Liemhetcharat and M. Veloso. 2014. Weighted synergy graphs for effective
over graphs. Journal of Artificial Intelligence Research (JAIR) 45 (2012), 165–196.
team formation with heterogeneous ad hoc agents. Artificial Intelligence 208
[32] A. Zha, K. Nomoto, S. Ueda, M. Koshimura, Y. Sakurai, and M. Yokoo. 2017.
(2014), 41–65.
Coalition structure generation for partition function games utilizing a concise
[17] N. Ohta, V. Conitzer, R. Ichimura, Y. Sakurai, A. Iwasaki, and M. Yokoo. 2009.
graphical representation. In International Conference on Principles and Practice of
Coalition structure generation utilizing compact characteristic function repre-
Multi-agent Systems (PRIMA). 143–159.
sentations. In International Conference on Principles and Practice of Constraint

1716