No TL;DR found
A pervasive assumption in Game Theory is that players’ utilities are concave, or at least quasi-concave, with respect to their own strategies. While mathematically instrumental, enabling the existence of many kinds of equilibria in many kinds of settings, (quasi-)concavity of payoffs is too restrictive an assumption. For the same reasons that (quasi-)concave utilities can only go so far in capturing single-agent optimization problems, they can only go so far in modeling the considerations of an agent in a strategic interaction. Besides, the study of games with non-concave utilities is increasingly coming to the fore as Deep Learning ventures into multi-agent learning applications. In this article, we study what types of equilibria exist in such games, and whether they are computationally tractable, proposing paths for Game Theory and multi-agent learning in the next one hundred years.