Using ℓp-norms for fairness in combinatorial optimisation
Introduction
The issue of fairness has received considerable attention from researchers in many fields, including, for example, computer science (Jain et al., 1984), economics (Nash, Jr, F., 1950, Rabin, 1993, Weymark, 1981), marketing (Xia et al., 2004), operational research (Bertsimas, Farias, Trichakis, 2011, Kelly, 1997), philosophy (Rawls, 1971), psychology (Janssen, 2000) and recreational mathematics (Dubins and Spanier, 1961). As one might expect, it has also received attention from the combinatorial optimisation community (see, e.g., Ajtai, Aspnes, Naor, Rabani, Schulman, Waarts, 1998, Barbati, Piccolo, 2016, Bertsimas, Farias, Trichakis, 2011, Fernández, Pozo, Puerto, 2014, Filippi, Ogryczak, Speranza, 2017, Karsu, Azizoglu, 2019, Karsu, Morton, 2015, Kostreva, Ogryczak, Wierzbicki, 2004, Luss, 1999, Martin, Ouelhadj, Smet, Özcan, Berghe, 2013, Matl, Hartl, Vidal, 2018, Muklason, Parkes, Özcan, McCollum, McMullan, 2017, Nguyen, Weng, 2017, Nickel, Puerto, 2005, Ogryczak, 2009, Puerto, Rodrıguez-Chıa, Tamir, 2017, Zukerman, Mammadov, Tan, Ouveysi, Andrew, 2008).
As noted in Filippi et al. (2017), many combinatorial optimisation problems (COPs) of interest can be modelled as follows. We have a set of workers and a set of tasks. Each task must be assigned to one worker, and the total cost depends on the allocation of tasks to workers. The issue of fairness then arises immediately, since, if one worker has a significantly higher workload than another, the solution may be perceived to be unfair.
A variety of approaches have been proposed to avoid unfair solutions, without incurring a significant increase in cost. (We survey these in Section 2.2.) In this paper, we focus on an approach suggested in Kostreva et al. (2004), that uses a modified objective function involving so-called ℓp-norms. We make the following specific contributions:
- •
We give an intuitive explanation for why the ℓp-norm approach tends to lead to fairer solutions.
- •
We present a generic formulation for ℓp-norm problems as convex mixed-integer nonlinear programs (MINLPs).
- •
For some specific COPs, we present alternative convex MINLP formulations that are much smaller, do not suffer from symmetry, and have a reasonably tight continuous relaxation.
- •
We give some encouraging computational results for some simple vehicle routing, facility location and network design problems.
The paper has a simple structure. The literature is reviewed in Section 2. The intuitive explanation and generic formulation are presented in Section 3. The specialised formulations are described in Section 4. The computational results are given in Section 5, and some concluding remarks are made in Section 6.
Section snippets
Literature review
We now review the relevant literature. Section 2.1 covers fairness in general, and Section 2.2 covers fairness in combinatorial optimisation.
General remarks on the ℓp-Norm approach
In this section, we make some general remarks on the ℓp-norm approach. In Section 3.1, we try to explain why unfair solutions arise in the first place, and why the approach can help. In Sections 3.2 and 3.3, we present generic MINLP and 0–1 LP formulations for COPs with ℓp-norm objective.
From now on, we say that the workers are identical if for all pairs of workers w, w′ and all sets S⊆T. In this case, we write c(S) instead of cw(S).
Compact formulations for specific problems
It turns out that, for certain specific COPs, one can obtain alternative MINLP formulations that are both small and free from symmetry. In Sections 4.1 to 4.3, we show this for the multiple TSP, the capacitated minimum spanning tree problem, and the k-median problem. Then, in Section 4.4, we discuss how to strengthen the continuous relaxations of the formulations.
Computational results
In this section, we present computational results for the m-TSP, CMSTP and k-median problem, for varying values of p. All MILP and MINLP formulations were solved to optimality using the mixed-integer LP and SOCP solvers of CPLEX1 12.7, using default settings. The experiments were run on a MacBook Pro with a 2.3 GHz Intel Core i5 processor and with 8GB memory. A time limit of 3600 s was imposed on each instance.
We remark that preliminary
Discussion
Fairness has received a great deal of attention in many disciplines. To achieve fairness in combinatorial optimisation, we recommend using the ℓp-norm approach. We have shown that, for certain network optimisation problems, it is possible to derive compact and convex MINLP formulations of the ℓp variants, that do not suffer from symmetry. We have also shown that outer-approximating the MINLP, using a relatively small number of perspective cuts, often enables one to obtain results of acceptable
CRediT authorship contribution statement
Tolga Bektaş: Methodology, Formal analysis, Software, Validation, Writing - original draft, Writing - review & editing. Adam N. Letchford: Conceptualization, Methodology, Formal analysis, Writing - original draft, Writing - review & editing.
Acknowledgements
We thank an anonymous reviewer for very useful comments, that have led to a significant improvement in the paper.
References (53)
- et al.
Second-order cone programming
Math. Program.
(2003) - et al.
Optimal vehicle routing with lower and upper bounds on route durations
Networks
(2015) - et al.
A state-of-the-art review of parallel-machine scheduling research
Eur. J. Oper. Res.
(1990) - et al.
Mixed integer nonlinear programming tools: an updated practical overview
Ann. Oper. Res.
(2013) - et al.
Beyond Average and Minimax in MILP
Technical report
(2017) Finding minimal center-median convex combinations (cent-dian) of a graph
Mgmt. Sci.
(1978)- et al.
An exact algorithm for the minimum squared load assignment problem
Comput. Oper. Res.
(2019) - et al.
Inequity averse optimization in operational research
Eur. J. Oper. Res.
(2015) Charging and rate control for elastic traffic
Trans. Emerg. Telecom. Technol.
(1997)- et al.
Location problems
Workload equity in vehicle routing problems: a survey and analysis
Transp. Sci.
The bargaining problem
Econometrica
Inequality measures and equitable locations
Ann. Oper. Res.
The price is unfair! a conceptual framework of price fairness perception
J. Market.
Approximation algorithms for minimum-load k-facility location
ACM Trans. Alg.
Fairness in scheduling
J. Algorithms
Solution of a min-max vehicle routing problem
INFORMS J. Comput.
Equality measures properties for location problems
Optim. Lett.
Branch-and-price: column generation for solving huge integer programs
Oper. Res.
The multiple traveling salesman problem: an overview of formulations and solution procedures
Omega
Data Networks
The price of fairness
Oper. Res.
Fairness and collaboration in network air traffic flow management: an optimization approach
Transp. Sci.
New techniques for cost sharing in combinatorial optimization games
Math. Program.
Worst-case analysis of a placement algorithm related to storage allocation
SIAM J. Comput.
Alternative formulations for the ordered weighted averaging objective
Infor. Proc. Lett.
Cited by (10)
Exploring fairness in food delivery routing and scheduling problems
2024, Expert Systems with ApplicationsIntra-facility equity in discrete and continuous p-facility location problems
2024, Computers and Operations ResearchA note on the exact solution of the minimum squared load assignment problem
2023, Computers and Operations ResearchGenerating balanced workload allocations in hospitals
2023, Operations Research for Health CareFairly Extreme: Minimizing Outages Equitably
2023, arXiv