Arena, A Distributed Asynchronous Reinforcement Learning package based on MXNet
Install the development version of MXNet first. After that, use python setup.py develop --user to install Arena.
Also, install the Arcade Learning Environment (ALE) if we want to run demo program that trains AI for Breakout.
To install MXNet extensions, using python mxnet-patch.py -p MXNET_PATH -t install. Uninstall these extensions by python mxnet-patch.py -p MXNET_PATH -t uninstall.
You need to recompile MXNet after installing the extensions. Also, the FFT/IFFT operators require CuFFT. Add -lcufft to the link flags in MXNet.
To compile & install MXNet on windows, the following cmake command may be helpful (run it under mxnet/build):
cmake -G "Visual Studio 12 2013 Win64" ^
-DCMAKE_BUILD_TYPE=Release ^
-DCMAKE_CONFIGURATION_TYPES="Release" ^
-DMKL_INCLUDE_DIR="C:\\Program Files (x86)\\IntelSWTools\\compilers_and_libraries_2016\\windows\\mkl\\include" ^
-DMKL_RT_LIBRARY="C:\\Program Files (x86)\\IntelSWTools\\compilers_and_libraries_2016\\windows\\mkl\\lib\\intel64\\mkl_rt.lib" ^
-DOpenCV_DIR="E:\ThirdParty\opencv-3.1\opencv\build\x64\vc12\lib" ^
-DZMQ_LIBRARY="D:\HKUST\libzmq\bin\x64\Release\v120\dynamic" ^
-DZMQ_INCLUDE_DIR="D:\HKUST\libzmq\include" ^
-DPROTOBUF_LIBRARY="D:\\HKUST\\protobuf\\cmake\\build\\release" ^
-DPROTOBUF_INCLUDE_DIR="D:\\HKUST\\protobuf\\src" ^
-DCUDNN_INCLUDE="C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v7.5\\include" ^
-DCUDNN_ROOT="C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v7.5\\lib\\x64" ^
-DBLAS=MKL ..
Keep parallelism in mind. Do not hide the key algorithms and keep the implementation readable and easy to hack.
Run python dqn_dist_demo.py to train AI for Breakout.