2048 is a very popular online game. It is very easy but hard to achieve its goal. Therefore we decided to develop an AI agent to solve the game. We explored two strategies in our project, one is ExpectiMax and the other is Deep Reinforcement Learning. In ExpectiMax strategy, we tried 4 different heuristic functions and combined them to improve the performance of this method. In deep reinforcement learning, we used sum of grid as reward and trained two hidden layers neural network. For ExpectiMax method, we could achieve 98% in 2048 with setting depth limit to 3. But we didn't achieve a good result in deep reinforcement learning method, the max tile we achieved is 512.
All the file should use python 3.5 to run.
Use the following code to install all packages.
pip install -r requirements.txtIf you want to use Chrome ctrl.
- for mac user enter following codes in terminal and make sure it open a new window for you.
/Applications/Google\ Chrome.app/Contents/MacOS/Google\ Chrome http://gabrielecirulli.github.io/2048/ --remote-debugging-port=9222- run
python 2048.py
The game infrastructure is used code from 2048-python.
The game contrl part code are used from 2048-ai
The class is in src\Expectimax\ExpectedMax.py.
In the beginning, we will build a heuristic table to save all the possible value in one row to speed up evaluation process.
In each state, it will call get_move to try different actions, and afterwards, it will call get_expected to put 2 or 4 in empty tile. Then depth +1 , it will call try_move in the next step.
The main class is in deep-reinforcement-learning.py.
The first version in just a draft, the second one use CNN as an architecture, and this method could achieve 1024, but its result actually not very depend on the predict result. The third version I implement a strategy that move action totally reply on the output of neural network. The result is not satsified, the highest score I achieve is only 512.
For more information, welcome to view my [report](AI for 2048 write up.pdf)