BlackjackAI

Machine Learning AI learns to play Blackjack

How to run

1. Run BlackJackAI.py

Configurations

1. Comment out line 140 to skip training and begin testing on pre-trained model of 1000 generations. Running line 140 will overwrite the model
2. Generations to train until stopping: line 132
3. Games to play after training: line 96
4. Set printing = True on line 4 in BlackJackSim.py to enable debug logging of cards.

Training

It seems assigning weights based on decisions at certain points (hit on 17, stand on 19, etc...) trains a smarter model than assigning weights based on wins. This is obvious because wins in blackjack can be circumstantial. The ideal point for the model to reach would be for it to recognize the dealer's shown card and factor that into it's decision. Of course, it's blackjack, so it's all luck anyway.

Consensus

The horrible sad truth to this neural network is that since there is luck involved, the model will always be skewed. A blackjack after a hit on 20 (although very unlikely) might teach the model that hitting on 20 is a good move, which it is not. To see moderate winrates (> 40%), the model has to run through quite a bit of games to understand how the game works.