Source code for the paper "Reinforcement Learning-Aided Design of Efficient Polarization Kernels"

The AlphaZero source code is based on minizero, please read the readme in minizero to setup the training enviroment.

The AlphaPolar environment can be found in minizero/environment/kernsearch folder, Currently, we didn't opensource our decoding algorithm (RMLD). You should replace the

complexity_ = 100; 

in minizero/environment/kernsearch.cpp with the complexity of your polar kernel decoder.
We might opensource the viterbi decoding algorihtm in the future.

To see the kernel generate by the AlphaPolar Agent, please setup the environemnt in minizero And run the following command

podman run --device nvidia.com/gpu=all  --cap-add=SYS_PTRACE --security-opt seccomp=unconfined \
  --network=host --ipc=host --rm -it -v .:/workspace kds285/minizero:latest

#train a 4x4 polar kernel 
tools/quick-run.sh train kernsearch gaz 50 -conf_str \
  env_board_size=4:actor_num_simulation=16:kern_search_game_len=20:network_type=conv:rand_init_step=0 -gen kernsearch_gaz.cfg
tools/quick-run.sh train kernsearch kernsearch_gaz.cfg 100

tools/quick-run.sh console kernsearch kernsearch_gaz_1bx256_n16-19e696-dirty_A16/
# showboard  
# genmove black
# gen_multi_move black
# ctrl + c

Note that we've modify kernsearch_gaz_1bx256_n16-19e696-dirty_A16/kernsearch_gaz_1bx256_n16-19e696-dirty.cfg
line 70

kern_search_game_len=1000

to make it work in console mode
but in training mode this number should be set to around 250, that is

#* ------16x16 board ----
tools/quick-run.sh train kernsearch gaz 50 -conf_str \
  env_board_size=16:actor_num_simulation=16:kern_search_game_len=250:network_type=conv:rand_init_step=7\
   -gen kernsearch_gaz.cfg
#* ------12x12board-----   
tools/quick-run.sh train kernsearch gaz 50 -conf_str   env_board_size=12:actor_num_simulation=16:kern_search_game_len=144\
:network_type=conv:rand_init_step=5 -gen kernsearch_gaz.cfg