Blackjack Reinforcement Learning

Overview

This project implements a Blackjack environment using Gymnasium and a PPO agent using Stable Baselines 3. You can customize the number of decks in the game by changing the --deck_size argument. Motivation: if we include seencards history in the state, can we imporve the odds.

Base Agent

State:

• Probabilities of each card in the deck.
• Entropy of the deck.
• Player's hand value.
• Dealer's visible card.

Actions:

• No Bet
• Bet
• Hit
• Stand
• Double

Rewards:

• Pealize for invalid actions.
• Penalize propotional to entory. To encourage the agent to bet more as we get more information (from seen cards)