Reinforcement learning framework for discovering Monte Carlo algorithm on ice model.
This is an official repository of paper "Generation of ice states through deep reinforcement learning "
General
- CMake (>= 2.8.3)
- Boost (>= 1.3.2)
- Python 2.7
- GCC
Python
- matlotlib
- Tesnorflow 1.4
For Mac OSX, we need to install extra boost-python library.
brew install cmake boost-python
for more details, please refer to https://github.com/TNG/boost-python-examples
- Compile icegame core
- Install gym-icegame interface (follow instructions in icegame2)
The inference should be executed at the folder rlloop. Go to the folder and download the trained model.
cd a3c
sh download.sh
or download model from https://drive.google.com/drive/folders/15MO-S_po4NIKsBL94rhOG5rC-fMbBn18?usp=sharing.
Now, we can play with it.
python play_icegame.py --log-dir saved_model
Use --render
for visualization.
The following command will launch 8 workers 1 parameter server and 1 rewards monitor.
python distribute_tasks.py -w 8 -l logs/my_task
For training, it takes about 3 days on 12 cpu cores.
The code in a3c_measure/ folder is modified to measure the correlation function, structure factor, probability frequency.
Download the Models and put in the folder to run.
sh run_*.sh