ZeroDivisionError: float division by zero

[mshanak]

I am trying to run this code on a customer game, The experiment works for some iterations but then but I keep getting the following error:

Traceback (most recent call last):
  File "main.py", line 61, in <module>
    main()
  File "main.py", line 57, in main
    c.learn()
  File ".../alpha-zero-general-master/Coach.py", line 89, in learn
    iterationTrainExamples += self.executeEpisode()
  File "... /alpha-zero-general-master/Coach.py", line 58, in executeEpisode
    pi = self.mcts.getActionProb(canonicalBoard, temp=temp)
  File ".... /alpha-zero-general-master/MCTS.py", line 56, in getActionProb
    probs = [x / counts_sum for x in counts]
  File "...alpha-zero-general-master/MCTS.py", line 56, in <listcomp>
    probs = [x / counts_sum for x in counts]
ZeroDivisionError: float division by zero

The configuration arguments in main.py are as follow:

args = dotdict({
    'numIters': 1000,
    'numEps': 100,              # Number of complete self-play games to simulate during a new iteration.
    'tempThreshold': 15,        #
    'updateThreshold': 0.6,     # During arena playoff, the new neural net will be accepted if threshold or more of games are won.
    'maxlenOfQueue': 200000,    # Number of game examples to train the neural networks.
    'numMCTSSims': 50,          # Number of games moves for MCTS to simulate.
    'arenaCompare': 40,         # Number of games to play during arena play to determine if new net will be accepted.
    'cpuct': 1,

    'checkpoint': './temp/',
    'load_model': False,
    'load_folder_file': ('/dev/models/8x100x50','best.pth.tar'),
    'numItersForTrainExamplesHistory': 20,

})

Is there any solution or explanation for counts_sum zero value?

[jrbuhl93]

I would guess self.game.getActionSize() is returning 0.

[mshanak]

I would guess self.game.getActionSize() is returning 0.

I printed the counts variable when the if counts_sum==0, the result was a vector of zeros [0,0,0..,0], this means that self.game.getActionSize() is returning the correct size.

I wonder why the following line returning an all-zero vector [0,0,0..,0]

        counts = [self.Number_visits_sa[(s, a)] if (s, a) in self.Number_visits_sa else 0 for a in range(self.game.getActionSize())]

[jrbuhl93]

That is unexpected since you have 'numMCTSSims': 50, so it should be performing 50 searches. Maybe you could put print statements where Nsa is incremented to make sure that is being called. Another possibility is that somehow the s and/or the a are not matching the values used when incrementing the count.

[goshawk22]

I am also having a similar problem. I am implementing a chess version and everything works when I pit two human players, or two random players, together, but the MCTS throws a Float Division by zero error.
I did some debugging and it seems that every value in counts is equal to zero. If temp = 0, the MCTS algorithm chooses randomly from counts
bestAs = np.array(np.argwhere(counts == np.max(counts))).flatten()
and, because not all of these moves are valid, the best action can be an illegal move and cause an error.
I'm not really sure where this problem is caused (there isn't an error when playing the game with two human players, and a win is correctly identified).

[goshawk22]

I worked out what was wrong with my implementation - don't know if it will help anyone else.
I wasn't copying the board before getting the next state which meant that whenever the MCTS algorithm explored an action, it would apply the action to the real board. A simple mistake!

[jamesbraza]

I am hitting this on the first invocation of MCTS.getActionProb with a chess.Board, where my getNextState could in-place mutate the board:

    a b c d e f g h
  . --------------- .
8 | r n b q k b n r | 8
7 | p p p p p p p p | 7
6 | . . . . . . . . | 6
5 | . . . . . . . . | 5
4 | . . . . . . . . | 4
3 | . . . . . . . . | 3
2 | P P P P P P P P | 2
1 | R N B Q K B N R | 1
  * --------------- *
    a b c d e f g h

After a few MCTS.search calls here in MCTS.getActionProb, you can see the effects:

    a b c d e f g h
  . --------------- .
8 | r n b q k b . r | 8
7 | p p p p p p p p | 7
6 | . . . . . . . n | 6
5 | . . . . . . . . | 5
4 | . . . . . . . . | 4
3 | . . . . . . . N | 3
2 | P P P P P P P P | 2
1 | R N B Q K B . R | 1
  * --------------- *
    a b c d e f g h

Thank you @goshawk22 for figuring this out! Imo it would be good to document in Game.py what methods are allowed to in-place mutate game state, and which are not. I will open a PR for this in the coming weeks.

[jamesbraza]

Hello all, as I have a few other PRs (#299, #300) that went unreviewed here, I am going to not contribute a PR for this.

For those hitting this in the future, feel free to reference my Game implementation here: https://github.com/jamesbraza/cs234-dreamchess/blob/main/azg_chess/game.py

• getValidMoves, getGameEnded, getSymmetries, stringRepresentation: doesn't mutate or copy the passed in board
  • Think: read only
• getNextState: doesn't mutate the passed in board. Instead will (1) make a copy of the canonicalized board, and (2) possibly mutate that copy
  • Think: read only, but can copy to write
• getCanonicalForm: can mutate the passed in board, for an optimization (keep reading)
  • Think: can write

The placement of read only is important because MCTS does store board state. At the same time, having getCanonicalForm be able to intelligently mutate the board can speed all future getCanonicalForm calls without affecting the stored board inside MCTS. Hopefully this helps someone, cheers!