Abstract: In this talk, I show how to apply statistical machine learning to design outcome rules in settings without money, looking to impose either good incentive or stability properties, while best approximating a target outcome rule. The approach leverages existing characterizations of families of mechanisms for problems of social choice and matching, and uses statistical machine learning to design a parametrized outcome rule that best optimizes other design objectives. Applications are demonstrated to strategy-proof social choice with single-peaked preferences, strategy-proof one-sided matching, and stable two-sided matching.

Based on "Learning Strategy-proof Mechanisms for Social Choice and Matching Problems," H. Narasimhan, S. Agarwal and D. C. Parkes, working paper, 2015.