Self-learning Black Jack player based on Reinforcement Learning

This repository contains a self-learning Black Jack player based on reinforcement learning.
The reinforcement learning methods Q-Learning and SARSA were implemented.
The Black Jack environment provided by OpenAI is used.
https://github.com/openai/gym/blob/master/gym/envs/toy_text/blackjack.py

The extensions are mostly based on
https://github.com/ml874/Blackjack--Reinforcement-Learning

The goal of this repository is to implement the following rules:

1. the Basic Strategy from [1]
2. the Complete Point-Count System from [1]
3. in addition to the basic rules, two rule variations
   • Rewards
     • Positive Reward times two
     • Negative Reward times two
   • Dealer stops
     • Dealer stops at 16
     • Dealer stops at 18

[1] Edward O. Thorp, "Beat the Dealer", Vintage, 1966

ToDo:

Splitting does not work correctly yet. Here there is still room for improvement.

Overview of the folder structure and files

Files	Description
Plots/	contains all tested hyperparameters and the corresponding images
agent.py	provides the agents for Q-Learning and SARSA
deck.py	provides the deck for card counting
environment.py	provides the environment
play.py	brings everything together here and can be seen as the starting point

Overview of the experiments and achieved results

Drawing & Standing is used in every attempt. The evaluation is based on the average payout. The corresponding image can be found in the Plots folder. A positive result is achieved only once.

Learning algorithm	Doubling Down & Splitting Pairs	Complete Point-Count System	Training epochs = Evaluation players	Evaluation epochs per player	Average Payout	Extension
SARSA	False	False	100	100	-18.02
SARSA	False	False	50000	1000	-186.29
Q-Learning	False	False	100	100	-30.58
Q-Learning	False	False	50000	1000	-63.45
Q-Learning	True	False	50000	1000	-109.55
Q-Learning	True	True	50000	1000	-103.94	Number of decks 6
Q-Learning	True	True	50000	1000	-85.10	Number of decks 4
Q-Learning	True	True	50000	1000	0.71	Number of decks 1
Q-Learning	True	False	50000	1000	-98.02	Positive reward * 2
Q-Learning	True	False	50000	1000	-269.54	Negative reward * 2
Q-Learning	True	True	50000	1000	-20.10	Positive reward * 2
Q-Learning	True	True	50000	1000	-313.13	Negative reward * 2
Q-Learning	True	False	50000	1000	-73.49	Dealer stops at 16
Q-Learning	True	False	50000	1000	-53.40	Dealer stops at 18
Q-Learning	True	True	50000	1000	-101.18	Dealer stops at 16
Q-Learning	True	True	50000	1000	-19.52	Dealer stops at 18