Zhiyang He, Charles Lu, Stephen Ou

CS 221 at Stanford University with Percy Liang, Autumn 2015

We explore various techniques to play 2048:

1. Randomized and pure strategies
2. Search strategies, i.e. expectimax and minimax
3. Reflex strategies, i.e. regression and neural networks
4. Reinforcement strategies, i.e. Q-learning

To run the front end:

1. Install Flask using pip install Flask
2. In the main directory, run python app.py, then navigate to localhost:5000

• gameState.py: provides an abstraction for the 2048 game
• agents.py: contains logic behind each of our AI agents
• play.py: backend only simulation entry point
• eval.py: implements various evaluation functions to score game board
• data.py: methods to read and write raw board and move label data for reflex model training
• tests.py: unit tests for the 2048 game state and evaluation functions
• util.py: miscellanous utility functions
• reflex.py: Theano script to train various reflex models
• appQ.py: Flask server used to connect Q-learning agents
• gameStateQ.lua: the Lua training script that appQ.py connects
• gameStateQ.py: backend-only Q-learning, does the same thing as appQ.py without frontend
• gameStateQInit.lua: Lua script used to reset local data files
• deepqlearn.{lua,moon}: DeepQLearning framework by blakeMilner
• app.py: the main entry point for Flask (a Python framework), used to connect search agents
• templates/index.html: homepage for the 2048 game frontend
• statics/*.js: logic for the 2048 game (forked from the original 2048 game by Gabriele Cirulli)