Signaling in Quasipolynomial time

TL;DR

This paper develops quasi-polynomial time algorithms for the signaling problem in game theory and auctions, using a meshing scheme to discretize belief spaces and optimize signaling schemes.

Contribution

It introduces a novel meshing scheme for discretizing belief spaces, enabling approximation algorithms for signaling in normal form games and auctions in quasi-polynomial time.

Findings

Algorithms run in quasi-polynomial time under various conditions.

Meshing scheme effectively overcomes the curse of dimensionality.

Convex partitioning and submodular maximization techniques are used for scheme assembly.

Abstract

Strategic interactions often take place in an environment rife with uncertainty. As a result, the equilibrium of a game is intimately related to the information available to its players. The \emph{signaling problem} abstracts the task faced by an informed "market maker", who must choose how to reveal information in order to effect a desirable equilibrium. In this paper, we consider two fundamental signaling problems: one for abstract normal form games, and the other for single item auctions. For the former, we consider an abstract class of objective functions which includes the social welfare and weighted combinations of players' utilities, and for the latter we restrict our attention to the social welfare objective and to signaling schemes which are constrained in the number of signals used. For both problems, we design approximation algorithms for the signaling problem which run in quasi-polynomial time under various conditions, extending and complementing the results of various recent works on the topic. Underlying each of our results is a "meshing scheme" which effectively overcomes the "curse of dimensionality" and discretizes the space of "essentially different" posterior beliefs -- in the sense of inducing "essentially different" equilibria. This is combined with an algorithm for optimally assembling a signaling scheme as a convex combination of such beliefs. For the normal form game setting, the meshing scheme leads to a convex partition of the space of posterior beliefs and this assembly procedure is reduced to a linear program, and in the auction setting the assembly procedure is reduced to submodular function maximization.