dependencies

install nvidia + cuda + cudnn

// linux disable secure boot, disable fast boot

sudo service gdm stop echo 'blacklist nouveau' > /etc/modprobe.d/blacklist-nvidia-nouveau.conf echo 'options nouveau modeset=0' >> /etc/modprobe.d/blacklist-nvidia-nouveau.conf

sudo apt-get install gcc make sudo dpkg -i cuda-repo-ubuntu2004-11-2-local_11.2.2-460.32.03-1_amd64.deb

// install repository

sudo apt-get install cuda sudo dpkg -i libcudnn8-dev_8.1.1.33-1+cuda11.2_amd64.deb libcudnn8_8.1.1.33-1+cuda11.2_amd64.deb

libraries

cd ~/Downloads
mkdir libtorch-cxx11-abi-shared-with-deps-1.8.1+cu111
unzip libtorch-cxx11-abi-shared-with-deps-1.8.1+cu111.zip
mv libtorch libtorch-cxx11-abi-shared-with-deps-1.8.1+cu111 

sudo apt-get install libopencv-dev libprotobuf-dev nlohmann-json3-dev libgtest-dev libboost-dev protobuf-compiler

build tensorboard_logger

cd third_party/tb_logger && make

remote install

apt-get install cmake vim wget unzip wget https://download.pytorch.org/libtorch/cu111/libtorch-cxx11-abi-shared-with-deps-1.8.1%2Bcu111.zip unzip libtorch*.zip apt-get install libopencv-dev libprotobuf-dev nlohmann-json3-dev libgtest-dev libboost-dev protobuf-compiler

article

• intro: jackal game
• existing algos: alphago, alphago-zero, mu-zero
• encoding: cnn vs lstm, arrows encoding, integer encoding, state/action encoding. Loss function
• performance: multithreading c++ vs python, bottlenecks of state copy/encoding, queue based model execution,
• training: replay buffer, model degrades, state function only?
• determinization: closed tiles

TODO

• test state value of synthetic game states