Skip to content

Simulate different methods Blackjack Dealers may be cheating, and train an adversarial model to take the edge back.

License

Notifications You must be signed in to change notification settings

kreshank/Unfair-Blackjack-Simulation

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

18 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Unfair Blackjack Simulations

In casinos, we can notice that blindly playing by the "book" often leads to catostrophic losses, much lower than the theoretical winning percentage. This repository provides two things: a game Environment (different settings of Blackjack, with options for non-random shuffling), and a player Agent which takes the place of a player.

Hypotheses

My personal hypothesis is that casinos often do a "rigged" shuffling that artifically groups aces and lower rank cards together, thus effectively eliminating the player's chance of getting a natural blackjack. Mathematically, this greatly reduces the player's odds by quite a significant margin. Some other possibilities (but less likely) is an unfair distribution of cards or burning cards frequently to either throw off player's count or alter card sequence.

This simulator allows for all of these settings to be tested and analytic data will be shown.

Notably with the TrainedPlayer class -- this is a trainable neural network that will learn 1) a card counting strategy, 2) betting strategy depending on running count, and 3) per-hand decision strategy. The card counting strategy can also be overridden with a fixed strategy if desired (Hi-Lo, K-O, etc.).

Methodology

Network Player

This is a standard neural network trained with Monte-Carlo REINFORCE and gradient clipping. The training is somewhat unstable as each epoch randomizes the setting: shoe size, number of players (other players play by Book), etc. For training, we can modify batch sizes and number of episodes for stability.

About

Simulate different methods Blackjack Dealers may be cheating, and train an adversarial model to take the edge back.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages