Playing 7 e mezzo with Deep Q-Learning

Inspired by the paper ”Playing Blackjack with Deep Q-Learning” by Allen Wu, I have developed this Reinforcement lerning project for a university exam.

Install and run

1. Create your own virtual env:
2. Install the dependencies in requirements.txt
3. Select the algorithm and the number of episodes in main.py
4. Run the main

Experimental results:

Policy evaluation

The first experiment was conducted to get an idea of the best policies, I evaluated each naive policy cause of the space of states isn't extremely large. The policies were chosen for each (player value, probability of bust) pair.

Example: Policy (4, 50)

If the player has in its hand a total value less than or equal to 4 and the probability of bust is less than or equal to 50% he takes a card otherwise he stay.

Below is a table with the winning percentage for each policy, the percentages have been calculated on 150,000 episodes.

We can see that the best results are obtained with policies whose player value limit is between 2.5 and 4 and the bust probability between 15% and 25%.

---

In following examples 0 stands for "HIT" and 1 stands for "STAY".

Q-Learning

Deep Q-Learning