2048-api

A 2048 game api for training supervised learning (imitation learning) or reinforcement learning agents

Statement

It is a course project and TA provides us with an expectimax agent which can get scores above 2048. Therefore, we can collect board and the agent’s corresponding solution (direction: up, down, left, right). It’s just like an image classification task where boards are parallel to images, solutions are parallel to labels. Our aim is to make our own agent's score reach 900+.

Deep learning part. My net: game2048/Model.py, training: game2048/train_CNN.py.

Code structure

• game2048/: the main package.
  • game.py: the core 2048 Game class.
  • agents.py: the Agent class with instances. Added LXYAgent
  • displays.py: the Display class with instances, to show the Game state.
  • expectimax/: a powerful ExpectiMax agent by here.
  • mergedata.py: preprocess the data obtained from ExpectiMax
  • Model.py: My Net
  • train_CNN.py: train my Net
  • para.pkl: the best parameters for my net
• explore.ipynb: introduce how to use the Agent, Display and Game.
• static/: frontend assets (based on Vue.js) for web app.
• get_layer_boards/:
  • agents.py: get data from Expectimax Agent
  • getboards.py: get data from Expectimax Agent
• webapp.py: run the web app (backend) demo.
• evaluate.py: evaluate LXYAgent.
• generate_fingerprint.py: get LXYAgent's fingerprint.
• past_train_log.out: part of my past train log.

Requirements

• code only tested on linux system (ubuntu 16.04)
• Python 3 (Anaconda 3.6.3 specifically) with numpy and flask
• Pytorch

To train the net

Before train the net, you should create the dir ./game2048/saved
then:
python3 train_CNN.py

To evaluate LXYAgent

python3 evaluate.py >>evaluation.log

To generate fingerprint

python3 generate_fingerprint.py

To compile the pre-defined ExpectiMax agent

cd game2048/expectimax
bash configure
make

To run the web app

python webapp.py

LICENSE

The code is under Apache-2.0 License.