How I Built a Pluribus-Style Poker AI From Scratch

Short summary

A technical tutorial on building poker AI using Counterfactual Regret Minimization (CFR), which solves imperfect information games by computing Nash equilibrium strategies. Covers Monte Carlo CFR for efficiency, hand abstraction via Earth Mover's Distance, Deep CFR neural networks to replace lookup tables, and real-time search refinement. Includes working code and practical optimizations like eval mode caching that improved speed 6x.

• •CFR algorithm finds Nash equilibrium by iteratively updating strategies based on regret from missed actions
• •Deep CFR uses neural networks instead of lookup tables to scale to complex games like No Limit Hold'em
• •Real-time search bootstraps from a blueprint strategy to refine decisions at runtime

Generated with AI, which can make mistakes.