blackjack

A (lame and unfinished) attempt at implementing the NEAT algorithm to learn to play soft 17 blackjack.

The goal was to have the AI learn to play blackjack in the most optimal way with as little influence valuing hands from me

The neural network has 3 inputs and 2 outputs

• Player hand value(4 - 20)
• Player soft ace present(0 or 1)
• Dealer value(2-11)
• Double decision(double or continue to next option)
• Hit decision(Hit or stand)

Unfinished elements

• Bet size output and wallet input has not been added. Also posibly a streak input to go along with this to see if bets would be increase/decreased during a winning/losing streak
• Split decision has not been added although the players hand is already kept in an array and the hand in played in a for loop that only runs once. Would be easy to add
• Threading

Issues

The outputs seem to follow a "linear" trajectory. The NN tends to either disregard to dealer value or not value it strong enough.

Examples:

• If the user has a 16 it will either hit or stand on all dealer cards. There are some cases where it develops to have different outcomes on a user card but it follows a 'diagonal' pattern in the chart that is printed. A playerr with 12 will stand until the dealer has 5. When the player has 13 it will stand until the dealer has 6. When a dealer has 14 it will stand until the dealer has 7 and so on.

• Weaknesses in the design are more apparent when doubling. When playing blackjack a user without an ace should only double when they have 9. 10 or 11 in their starting hand. The 'linear' result can never double a 10 and not double an 8.

• The only potential fix to this I can think of would be to add more inputs to better represent the 'non-linear' odds of blackjack. I've been avoiding that to not just make the neural network remember every decision individually but the current input/output doesn't work.

Statistics class is no longer used in the initial way it was intended to.

• The class was meant to keep the best players stats and the overall stats
• Currently only keeps overall stats but best player stats are still implemented
• Could be changed to save every individuals stats