---
tags:
- reinforcement-learning
- blackjack
- q-learning
- tabular
license: mit
---

# Blackjack Q-table Policy

This repository contains the trained tabular Q-learning policy used in the IEMS5726 Blackjack Reinforcement Learning Trainer project.

## Model

- Algorithm: tabular Q-learning
- Policy artifact: `q_model.json`
- Actions: `0 = stand`, `1 = hit`, `2 = double_down`
- State representation: `(player_total, dealer_upcard, usable_ace, can_double, true_count_bucket)`
- Rules: one player versus dealer, finite 6-deck shoe, Hi-Lo running count, true-count bucket, double down, dealer stands on soft 17, natural Blackjack pays 3:2.

## Evaluation

The selected policy was evaluated over 1,000,000 Blackjack hands.

| Metric | Value |
| --- | ---: |
| Average reward | -0.0086975 |
| Win rate | 0.433611 |
| Loss rate | 0.481673 |
| Draw rate | 0.084716 |

`q_learning_model_comparison.csv` compares the selected expert-prior/count policy against 5,000,000-hand fine-tuning variants.

## Files

- `q_model.json`: trained Q-table policy and evaluation metadata
- `policy_table.csv`: exported policy table
- `policy_heatmap.svg`: policy visualization
- `training_history.csv`: training-history file generated by the training pipeline
- `q_learning_model_comparison.csv`: final model comparison table

## Usage

The browser demo can load this model JSON directly from the public URL. In the project submission, place the public link in `application/model_link.txt`.