Skip to content

Latest commit

 

History

History
36 lines (21 loc) · 1.66 KB

README.md

File metadata and controls

36 lines (21 loc) · 1.66 KB

Reinforcement Learning

Requirements

  • Be sure that you had built the engine. The engine should be in the build directory. Recommend to use the -DUSE_ZLIB=1 option to save your disk memory.
  • PyTorch 1.x (for python)
  • NumPy (for python)

Performance

Use the default setting in the bash directory. The network will at least reach strong amateur level in 1 ~ 2 weeks on 19x19 with a RTX 2070s computer.

Simple Usage

There are two bash files. The setup.sh will copy the engine and training script to the current directory. Then executing simple.sh and start simple synchronized loop. And default will use all GPUs.

$ cp -r bash selfplay-course
$ cd selfplay-course
$ bash setup.sh -s ..
$ bash simple.sh

The simple.sh will do the infinite loop. If you want to halt the loop, you need to create a kill.txt file and wait for end of this round.

$ touch kill.txt

Maybe you have a powerful computer with multi-GPUs but don't want to use all GPUs. You may add the option -g to execute the self-play and training on the specific GPU.

$ bash simple.sh -g 1    # run it on the 2nd GPU

Or select specific multi-GPUs.

$ bash simple.sh -g 1 -g 2    # run it on the 2nd and 3rd GPU

Sample Configuration File

The sample directory includes some enigne self-play Configuration files. The sample/full-gumbel-p16.txt will do full Gumbel learning with 16 playouts. The sample/full-alphazero-p400.txt will do full AlphaZero learning with 400 playouts. You may simply reuse the files for customization learning. And please see this section. It explains the configuration and training parameters.