codygillespie / 2048-reinforcement-learning

A simple implementation of 2048 and DQN agent to learn the game

2048 Reinforcement Learning Agent

Introduction

This project is a reinforcement learning agent that learns to play the game 2048. The agent is trained using a modified version of the game that allows the agent to play the game without any human interaction (see game.py).

The approach used is Deep Q-Learning with Experience Replay. The agent is trained using a simple fully connected neural network with 2 hidden layers (deep.py). The input to the network is the current state of the board, and the output is the Q-value for each possible action.

game.py

You can play the game directly by running python game.py. The game is played using the WASD keys.

main.py

Training can be started by running python main.py.