naughtsandcrosses not playing optimally about mcts

[WhymustIhaveaname]

I will have a look at it soon.

[WhymustIhaveaname]

I know it is impolite to say, but I have to say, it works on my machine. I tested it twice. Firstly, I let it play with itself.

searcher=mcts(timeLimit=1000)
gamestate=NaughtsAndCrossesState()
while not gamestate.isTerminal():
    action=searcher.search(initialState=gamestate)
    print("player %2d played %s"%(gamestate.getCurrentPlayer(),action))
    gamestate=gamestate.takeAction(action)
print("game over, result: %s"%(gamestate.getReward()))

Then, I test it with some must-win scenarios.

searcher=mcts(timeLimit=1000)
gamestate=NaughtsAndCrossesState()
gamestate.board=[[1,1,0],[-1,-1,0],[0,0,0]]
action=searcher.search(initialState=gamestate)
print("player %2d played %s"%(gamestate.getCurrentPlayer(),action))
gamestate=gamestate.takeAction(action)
assert gamestate.isTerminal()
print("result: %s"%(gamestate.getReward()))

It always gives me the best choice. Can you share at which board situation it breaks?

[pbsinclair42]

Can you also confirm which version of the package you were using when you found the above? I'm also unable to reproduce this I'm afraid.

[pbsinclair42]

Closing this due to no response, feel free to comment if this issue persists.

[acmol]

reproduce this too.
I install this lib with pip install, then download the demo naughtsandcrossses, then reproduced this bug.

Name: mcts
Version: 1.0.4
Summary: A simple package to allow users to run Monte Carlo Tree Search on any perfect information domain
Home-page: https://github.com/pbsinclair42/MCTS
Author: Paul Sinclair
Author-email: [email protected]
License: MIT
Location: /usr/local/lib/python3.8/dist-packages
Requires:
Required-by:

[acmol]

I found the reason.

The code installed by "pip install" has a bug fixed in master branch.

#10