Most real-world settings are imperfect-information games. They present challenges beyond those in perfect-information games. In 2017, our AI Libratus beat top humans in the main benchmark, heads-up no-limit Texas hold’em. In this talk I will discuss some of our more recent work on imperfect-information games. Topics include a unified framework for abstracting games with bounds on solution quality [Kroer & Sandholm, NeurIPS-18], a sound depth-limited search framework [Brown et al., NeurIPS-18], the fastest equilibrium-finding algorithms [Brown & Sandholm, AAAI-19], deep learning as an alternative to abstraction [Brown et al., Deep RL Workshop-18], a general framework for online convex optimization for sequential decision processes and extensive-form games [Farina et al., AAAI-19], the first scalable algorithm for trembling-hand equilibrium refinements [Farina et al., NeurIPS-18], and trembling-hand refinement of Stackelberg equilibria [Farina et al., IJCAI-18; Marchesi et al. AAAI-19].
