Reducing Bias in Preference Aggregation for Multiagent Soft Constraint Problems

Alexander Schiendorfer and Wolfgang Reif

Most distributed constraint optimization problems assume the overall objective function to be the ``utilitarian social welfare'', i.e., a sum of several utility functions, belonging to different agents. This also holds for the most popular soft constraint formalisms, cost function networks and weighted constraints. While, in theory, this model is sound, it is susceptible to manipulation and resulting bias in practice. Even without malevolent intentions, bias can result from the way orderings over solutions are transformed into numerical values or normalized. Alternatively, preferences can be aggregated directly using the tools of social choice theory to discourage manipulations and practically reduce unwanted bias. Several common voting functions can be implemented on top of constraint modeling languages through incremental search and suitable improvement predicates. We demonstrate that our approach, in particular Condorcet voting, can undo bias which is shown on two real-life-inspired case studies using the soft constraint extension MiniBrass on top of MiniZinc.
published 2019 Proceedings Principles and Practice of Constraint Programming. CP 2019

Publisher: Springer

DOI: Doi

For questions regarding the publication, please contact!