code.talks 2019 - The Building Blocks of Superhuman Poker AI
Автор: code.talks (ehem. Developer Conference)
Загружено: 2019-10-29
Просмотров: 4673
by Max Pumperla
Imperfect information games deal with taking strategic decisions when facing hidden information. Poker has been a classic benchmark for these games since the early days of AI. Games like Go and Chess are perfect information games, so the techniques used to solve them don't work for imperfect information games. For instance, poker-solving algorithms need to take into account how opponents might adapt to and exploit your strategy, which makes it difficult to estimate the value of your cards during play. In this talk we introduce some of the core ideas used by the superhuman poker AI called Pluribus. This AI system uses a clever self-play algorithm based on counterfactual regret minimization (CFR) to compute a blueprint poker strategy. During actual game-play against humans it then improves its strategy by searching for better options in real-time. We give a glimpse at how to implement the basics of Pluribus in Python.
Доступные форматы для скачивания:
Скачать видео mp4
-
Информация по загрузке:
