Blackjack-RL: A Reinforcement Learning Exercise for Blackjack

Phase 0

Test against various baseline decision models as benchmarks

• dealer's own decision model
• random decision model
• a standard 'fixed' decision model based on dealer's showing card, player's total, and soft ace count
• a standard 'fixed' decision model like the aforementioned that also takes into account the hi-lo card count

In all of the above, the bet amount is fixed (never varies, player always bets the minimum)

Phase 1

Test against a Q-learning trained model that does not count cards, and the bet amount is fixed. Its output is only hit/stand.

Phase 2

Test against a Q-learning trained model that takes the hi-lo card count into account, and the bet amount is fixed. Its output is only hit/stand.

-- we are here --

Phase 3

Test against a Q-learning trained model that takes the hi-lo card count into account, and varies its bet amount accordingly. The model for choosing hit/stand is the same model from Phase 2, and choosing an optimal bet amount will use a new model which will be trained with usage of an already-trained instance of the Phase 2 model.