hive-with-ai-prolog

Implemention in prolog of known board game Hive. We don't see ourselves in the future improving the project or fixing bugs, but feel free to make all the improvements you want or contacting us for help.

Using SWI-Prolog

cd gui
swipl main.pl

Currently File > New Game > Local game works out of the box.

To Start Up AI

Create a db folder inside ai (yes, it is ugly). Then you can play against a dumb IA by starting a New Game in the GUI and selecting against Ai or AI vs AI.

AI uses Monte Carlo Search Trees (MCSTs) and it is currently untrained so it will be dumb as f*ck.

To train the AI

cd gui
python run_train.py <times>

where <times> represents the amount of matches the IA is going to play against itself. All nodes are stored persistently ai/db/tree_db.pl.

When the AI has found an average of 2000 ~ 4000 leaf nodes it start doing some good looking stuff, although not enough.

Possible Improvements

• (very easy) Improve game art (which is already awesome btw) by modifying/changing bug pictures in gui/graphics/xpm. Replace by files with the same name. Images must be 70x70 pxs. Be careful new bugs are not bigger than the game hexagon.
• (easy) Add logic to to detect end game states in gui. A solution can be to obtain current game status by a call to gui_api and if it is not non_terminal do not accept any more events.
• (easy) Correctly move game\gui_api.pl to gui\gui_api.pl and update code properly so no errors occurs. gui_api.pl is out of place right now.
• (easy) Add logic to automatically create a db folder in ai when this folder does not exists.
• (medium) Make possible restarting a game without needing to exit current process.
• (hard) Make MCST multithreading. It will need to use mutexes to avoid race conditions when writing to ai/db/tree_db.pl. It will also need to make global game status local to each thread, since simulations are run over this by changing it, and setting it back to the initial state when finished.
• (hard) Add an efficient way of storing the MCST nodes. Currently persistent library is used, when around 4000 branches have been explored the database weights around 80 mbs, and takes around 3 seconds to load.
• Any other thing you think is cool.