Continuous Blackjack

Refactored simulation framework for continuous blackjack strategy research.

Game Rules

1. Players are randomly re-ordered every round.
2. On a turn, a player repeatedly draws U(0, 1) and accumulates the sum.
3. A player stops according to its strategy.
4. A score is the stopping sum if it is < 1, otherwise 0.
5. Highest score wins the round (ties go to the earliest max in current implementation).

Project Layout

continuous_blackjack/
  core/          # engine, round record, strategy base classes
  strategies/    # equilibrium, adaptive, statistical, bandits
  rl/            # optional torch-based DQN and actor-critic
experiments/
  run_experiment.py
tests/

Install

Python requirement: 3.14+

python -m pip install -e .

Optional extras:

python -m pip install -e ".[dev,analysis,rl]"

Or with requirements files:

python -m pip install -r requirements.txt
python -m pip install -r requirements-analysis.txt
python -m pip install -r requirements-rl.txt
python -m pip install -r requirements-dev.txt

Run Example Simulation

python experiments/run_experiment.py --set-index 1 --blocks 20 --rounds-per-block 1000 --log

Actor-Critic experiment:

python experiments/run_experiment.py --actor-critic --blocks 20 --rounds-per-block 1000 --log
# or:
python experiments/run_actor_critic_experiment.py --blocks 20 --rounds-per-block 1000 --log

Save/load RL model parameters (DQN / Actor-Critic):

python experiments/run_experiment.py --actor-critic --blocks 20 --rounds-per-block 1000 --save-model-dir checkpoints/ac
python experiments/run_experiment.py --actor-critic --blocks 20 --rounds-per-block 1000 --load-model-dir checkpoints/ac --save-model-dir checkpoints/ac

Notebook:

• /Users/muzhao/Documents/Workspace/Python/Continuous-Blackjack/experiments/continuous_blackjack_experiment.ipynb
• Configure load_model_dir / save_model_dir in the parameter cell to resume RL models.

Run Tests

python -m pytest

Notes

• No backward-compatibility layer is provided in this refactor.
• RL modules require torch; core and non-RL strategies do not.