2048 Machine Learning Bot

An advanced, high-performance Reinforcement Learning agent designed to autonomously play and master the game 2048. Utilizing Temporal Difference Learning (TD(0)) and Symmetric N-Tuple Networks, this bot is capable of achieving superhuman scores.

Overview

This project implements a headless training environment to build a massive 16.7 million state neural network, alongside three distinct browser automation scripts to deploy the trained agent in real-time. The system evaluates the board across all 8 mathematical symmetries simultaneously, allowing it to learn at an accelerated rate.

Features

• N-Tuple TD(0) Learning: Evaluates complex 6-tile shapes (rectangles and L-shapes) to accurately predict the value of any board state.
• Symmetric Sampling: Automatically applies learned patterns across all rotations and reflections.
• Expectimax Search: Looks multiple moves ahead, dynamically adjusting search depth based on the number of empty tiles to maintain high execution speed.
• Numba Acceleration: The training core is fully JIT-compiled and multithreaded, capable of playing thousands of games per second across all CPU cores.
• Atomic DOM Parsing: Uses injected JavaScript to instantly parse the browser DOM, eliminating race conditions during tile animations.

Execution Modes

The project is split into modular scripts to accommodate different use cases:

• train.py: The multithreaded training environment. Run this to train the model. The weights are continuously saved to weights.npy. You can adjust the speed using --speed 1/2/3.
• normal_bot.py: The classic experience. Plays exactly one game and waits for user input before closing the browser.
• bot.py: Target mode. Automatically clicks "Try Again" and loops infinitely until a specific target tile is reached. Defaults to 8192, but can be customized (e.g., python bot.py 16384).
• streamer_bot.py: An infinite loop designed for livestreams. Employs a persistent MutationObserver to permanently block ads and banners.

Setup and Installation

1. Clone the repository.
2. Install the required dependencies:

   pip install numpy numba playwright
   playwright install chromium

3. Run python train.py --speed 3 to begin building the neural network weights.
4. Once satisfied with the training progress, execute any of the bot scripts to watch the agent play.

Technical Architecture

For an in-depth breakdown of the mathematics and machine learning principles driving this bot, please refer to the ML_ARCHITECTURE.md file included in this repository.

Credits

Developed by Krshs90