Learning Correlated Equilibria in Extensive Form Games

Abstract

We formalize an efficient class of counterfactual regret minimization algorithms exploiting the “sequence form” to compute ‘‘$\Phi$-equilibria’’ – a generalized class of equilibrium concepts defined within general-sum extensive form games of imperfect information. We develop increasingly strong notions of no-regret, mapping those notions directly onto concepts of interest, such as agent- and extensive-form correlated and coarse correlated equilibria.