Learning tree structured potential games

Vikas K. Garg, Tommi Jaakkola

Research output: Contribution to journalConference articleScientificpeer-review

9 Citations (Scopus)

Abstract

Many real phenomena, including behaviors, involve strategic interactions that can be learned from data. We focus on learning tree structured potential games where equilibria are represented by local maxima of an underlying potential function. We cast the learning problem within a max margin setting and show that the problem is NP-hard even when the strategic interactions form a tree. We develop a variant of dual decomposition to estimate the underlying game and demonstrate with synthetic and real decision/voting data that the game theoretic perspective (carving out local maxima) enables meaningful recovery.

Original languageEnglish
Pages (from-to)1560-1568
Number of pages9
JournalADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS
Publication statusPublished - 2016
MoE publication typeA4 Article in a conference publication
EventIEEE Conference on Neural Information Processing Systems - Barcelona, Spain
Duration: 5 Dec 201610 Dec 2016
Conference number: 30

Fingerprint

Dive into the research topics of 'Learning tree structured potential games'. Together they form a unique fingerprint.

Cite this