Blackjack Reinforcement Learning

I built a Blackjack-playing AI that learns entirely through experience, training itself by playing thousands of simulated games. Instead of being programmed with fixed rules, the agent gradually discovers when to hit or stick by observing which choices lead to better outcomes over time. After enough practice, it develops strategy patterns that closely match optimal Blackjack play, even in tricky situations like dealing with strong dealer cards or managing soft hands. In the end, the agent reaches a stable win rate near the theoretical best, making the project a fun and rewarding demonstration of how reinforcement learning can teach an AI to make smart decisions.

Overview

AI agent plays Blackjack through self learning
Learns strategy by trying thousands of simulations
No fixed rules, discovers behavior independently

Implementation

Uses Q learning to evaluate state action values
Updates the Q table after each outcome
Learns when to hit or stick based on long term rewards

Results

Reached a win rate close to theoretical optimal
Handles soft hands and strong dealer cards better
Behavior becomes stable after enough training episodes

What I Learned

Reinforcement learning fundamentals
Exploration versus exploitation tuning
Analyzing training patterns and agent performance

Technologies Used:

Python, Gymnasium (Blackjack-v1 environment)