1. Remove the '-9' option from the CXX flags option in the makefile if your system does not support it. (The code runs perfectly fine with g++ alone on the Zoo) Everything should work after unzipping except one change MUST be made to the file features.hpp.

Navigate to /cpp and change the /MY/ABSOLUTE/PATH portion of the variables lmcache and train_in at the top of the file to the absolute path of wherever the project file has been unzipped. If you try to run and later on get some error message about json not parsing, it is because the paths here are incorrect.

• I did not quite have time to make a clean work around for this, but basically here we are setting the path to the json file that stores the cache of legal moves and the list of labeled training data. For some reason the json C++ package I am using complains if it does not have the absolute path.

1. Navigate to /cpp and run make. (By default this makes the shared library shogilib.so that gets used by python).
2. Go back to the main directory and activate the python virtual environment with: . venv/bin/activate
3. Run python3 evolve.py

Take a look at log.txt to see the results of the execution. Or since execution can take awhile even for 500 training positions, you can periodically check log.txt to see the progress of the GA. Note that the min and max values represent the square of the total number of positions guessed correctly.

NOTE : If you are having import errors there could be something wrong with the python venv. If that is the case, I have included a list of the dependancies I have successfully installed on the zoo. This is contained in requirements.txt and to setup a new venv this way run: python3 -m venv venv in the shogi directory, activate the venv again, then run pip install -r requirements.txt

• The main evolve.py script should print out all of the default GA settings / configuration for the algorithm, but these can very easily be changed at the top of config.py. They should be fairly self explanetory.

• By default there is already a loaded selection of 500 training moves in the the json directory and evolve.py is set to run the C++ version of evaluation with a very small population size. A larger sample of training data can be generated with sample.py (The paper used a size N=10,000 split in half between training and testing. See other options for details on how to run this script).

• Getting a larger sample size for training data. The script sample.py has what I think is a detailed enough arg parser that you can figure out what parameters to provide. However, be careful with the name and path of the in and out files. A correct path to the infile is necessary and right now both the cpp and python versoin of evolve funtion expect the cache file to have the name legal_moves_cache. The labeling formats also differ depending on if you are running cpp version or python.Here are the commands I used to generate the current set of 500 training positions for both the python and cpp versions.

  • CPP: python3 sample.py --fin data/games_small.txt --out legal_moves_cache --method sr --n 1000 --type json --seed 64 --hex True --move int
  • Python: python3 sample.py --fin data/games_small.txt --out legal_moves_cache --method sr --n 1000 --type pickle --seed 64
  • Lastly, depending on the size of the game file, sampling can take quite some time which is why I have just been working with /data/games_small which contains about 1000 games. I did not include my entire database file of over 60,000 games but can zip and send it over if you would like. It is also quite time consuming since I figured out pretty late that a surprising amount of the moves I collected from online are invalid. There are several typos, especially in games that were sampled from many years ago as they were inputted by hand I presume. Thus I have to additoinally check if a selected move is legal, which takes time with python-shogi. But again this only really needs to be done once and I have larger cache files saved on my machine already,

• There are some other options for make　too, mainly featuresTest which will run the main() from features.cpp. There are some comented lines of code in the evaluate_organism function that output the time it took for an individual organism evaluation that can be interesting to take a look at and compare to python.

  • Note: MUST remove -fPIC -shared from CXXFLAGS for this to work.

• If you are playing around with sample.py and are selecting new training data / relabeling it, by default it saves in an ugly format. Run make lmcache-pretty and go into lmcache.cpp and uncomment main() to save a more human readable copy of legal moves cache / training data to take a look at.

  • Note: MUST remove -fPIC -shared from CXXFLAGS for this to work.

• Take a look at features.py to see how I implemented some heuristic analysis. The code is quite commented and should be fairly easy to understand, but the python version is not very effecient. features.cpp implements basically all of these features with a new shogi API except for the last couple of the board shape.
• Of course, evolve.py has quite a bit going on as well, but here too it is commented well so it should be relatively easy to follow. There are some funky design choices about global definitons, but this is mostly due to some things I could not controll about the packages I was using / multiprocessing. But see comments for further details.
• The other one-off scripts I used to borrow and clean up my database of games is in /screipts

• Python Libraries:
  • DEAP
  • Pyhthon Shogi
  • Cpp Shogi
  • Open MP
  • Open GA