2048-Auto

2048 is a well known puzzle game, where the player will move number tiles in a 4×4 grid to merge the number 2048, or even higher.

The game logic is implemented in Python 3, and the user interface is implemented in HTML, CSS, and JavaScript. A small Flask server (app.py) exposes HTTP endpoints so the same environment can drive both the browser UI and automated agents.

Local training with Q-learning

The repository now ships with a lightweight Q-learning trainer that plays directly against the Python environment. It collects experience, updates a tabular policy, and can optionally resume from a saved checkpoint.

Run a training session (this also evaluates the greedy policy once training finishes):

python train_agent.py --episodes 5000 --log-interval 250

Key options:

• --episodes – number of training games to run.
• --alpha, --gamma, --epsilon, --epsilon-decay, --epsilon-min – standard Q-learning hyperparameters.
• --load-path – path to an existing bin/q_table.pkl to continue training.
• --save-path – where to store the trained table (defaults to bin/q_table.pkl).
• --eval-episodes – how many greedy runs to average for the evaluation summary.

After training you can point the front-end’s auto-play logic to the resulting policy or use the saved table to bootstrap more advanced agents.

Optional Postgres metrics logging

If you want to log per-episode metrics to Postgres (for visual inspection in tools like TablePlus), run Postgres with Docker and pass a connection URL to the training scripts.

docker compose up -d
export POSTGRES_URL="postgresql://2048_user:2048_pass@localhost:5432/2048_db"
python CNN/cnn_agent.py --episodes 1000 --db-run-name "cnn-run-1"
python MLP/mlp_agent.py --episodes 1000 --db-run-name "mlp-run-1"

Schema reference lives at db/schema.sql.