Policy network penalized incorrectly on invalid moves

[Timeroot]

Currently, the policy network is actively trained to output 0 on a move whenever it's invalid. The move is never taken, so target_pi's is zero there, and this enters into the loss function as a result. As a result, the policy network will be saying as much about "How often is this move legal" as it does about "Is this a good move" -- which is almost certainly hurting performance.

The correct action would be to manually mask out the result in NNet.py, right after getting the output from OthelloNNet.py: OthelloNNet.py returns F.log_softmax(pis), and then the entries there should be manually set to zero if it's not a valid move. This will prevent gradients from propagating back there.

[Timeroot]

Err, a correction: it should be masked with valid moves, and then re-normalized. Otherwise it's not changing anything.

[Timeroot]

For reference: I fixed this bug in my local copy, and it appears to be learning several times faster . Passing in the valid moves requires adapting a lot of the interfaces though, so, changing all the game implementations to do the masking. And my local copy is a huge mess, so I don't have anything I can easily offer as a PR.

But as some reference as to what I did, in Coach.py, the trainingExamples need to be extended to include what the valid moves are, so that's got something like

`

        pi = self.best_mcts.getActionProb(canonicalBoard, temp=temp)
        valids = self.game.getValidMoves(canonicalBoard, 1)
        bs, ps = zip(*self.game.getSymmetries(canonicalBoard, pi))
        _, valids_sym = zip(*self.game.getSymmetries(canonicalBoard, valids))
        sym = zip(bs,ps,valids_sym)

        for b,p,valid in sym:
            trainExamples.append([b, self.curPlayer, p, valid])

        action = np.random.choice(len(pi), p=pi)
        board, self.curPlayer = self.game.getNextState(board, self.curPlayer, action)

        r = self.game.getGameEnded(board, self.curPlayer)

        if r!=0:
            return [(x[0],x[2],r*((-1)**(x[1]!=self.curPlayer)),x[4]) for x in trainExamples]`

whereas obviously the ideal thing would be adapting getSymmetries to also permute valids, instead f calling it twice like I do.

Then MCTS.py needs to be changed to pass the valid moves in through nnet.predict, as in

valids = self.game.getValidMoves(canonicalBoard, 1) self.Ps[s], v = self.nnet.predict((canonicalBoard,valids))

Then NNet.py needs to have valids added a value to unpack, in both predict and train, and once prepped this should get passed in to the net -- so changing self.nnet(boards) to self.nnet((boards, valids)).

Finally, OthelloNNet.py needs to be changed to use the valids as a mask. So forward will start

def forward(self, boardsAndValids): s, valids = boardsAndValids

and include pi = valids * pi after computing pi. Obviously NNet.py and OthelloNNet.py need to appropriately changed in all libraries' implementations, for all games; and if you don't want to do the ugly getSymmetries hack like above, then that means also changing the convention for that in OthelloGame.py and all corresponding game definitions.

[Marcin1960]

"For reference: I fixed this bug in my local copy, and it appears to be learning several times faster ." @Timeroot

Wow!

[Timeroot]

I realized that my wording above was somewhat vague. pi = valids * pi works, but then you can no longer use log_softmax, since that has the normalization inside of it. So you'll likely need to sacrifice the numerical stability of log_softmax (which is probably fine), and replace

pi = log_softmax(pi)

with

pi = torch.exp(pi) pi = valids * pi pi = pi / torch.sum(pi)

The alternative is:

pi -= (1-valids)*1000 pi = log_softmax(pi)

which just reduces the log-probabilities to such a negative value as to be effectively zero.

[keesdejong]

@Timeroot Interesting. Although I am an absolute beginner, I already had suspicions in the same direction. I am going to try and implement your ideas in my own game for which I have largely used this git. I am curious about the result. It would be useful if you made the code that you created in one way or another accessible.

[keesdejong]

@Timeroot

I'm almost there, but I have a problem with implementing your idea in tensorflow. Maybe you can help me. I have a problem with this piece of your description:

replace pi = log_softmax (pi) with pi = torch.exp (pi) pi = valids * pi pi = pi / torch.sum (pi) The alternative is: pi - = (1-valids) * 1000 pi = log_softmax (pi)

In tensorflow.OthelloNet this should be around line 36: self.pi = Dense (s_fc2, self.action_size) self.prob = tf.nn.softmax (self.pi) self.v = Tanh (Dense (s_fc2, 1)) self.calculate_loss ()

I am new in this matter. Can you tell me how to do this for Tensorflow?

[evg-tyurin]

@keesdejong I'm looking forward for the confirmation that this trick can speed up the learning. I have my own implementation of the game of checkers here https://github.com/evg-tyurin/alpha-nagibator It's based on this repo so I could implement changes/make PR to both repos in the case it really helps.

[aletote]

Any news on when this is going to be updated for the main branch? Also, how much faster is it?

[51616]

I followed your fix and got small negative pi loss (around -0.00xx). Is this normal ? edit. now i'm using alternative code and it produces positive pi loss.

[jl1990]

@51616 If you made a fix, please upload PR

[51616]

@jl1990 i use this equation instead. pi -= (1-valids)*1000 pi = log_softmax(pi) this should produce positive pi loss.

[rlronan]

I made a pull with the changes introduced for Othello's Tensorflow implementation. Can someone confirm that these changes are correct? I can add them to the remainder of the games/implementations if that is the case,. As a warning, these changes are not backwards compatible with the saved models in the repository.

Also can someone provide some guidance as to how I would properly cite Timeroot for the changes, since I essentially just implemented what he suggested verbatim?

Here's the pull:

https://github.com/suragnair/alpha-zero-general/pull/163

[keesdejong]

@rlronan Thank you. I'll see if I find the time to check if it works. And, most importantly, whether it has any effect.

[mha-py]

This thread is a bit older, but since its pinned here is an idea: What about putting the information, if a move is invalid, into the p array of the replay list, with a value like -0.001 (like its done with draws)? Then the symmetry functions dont have to be changed (since it works with p) and its downward compatible: the value is close to zero, so the unupdated files for the games will still work.

[cestpasphoto]

Adding my opinion on this (old) thread: I implemented the formula proposed above pi -= (1-valids)*1000 I also implemented this paper (check also the related github repo) with pi = torch.where(valids, pi, -1e8)

The second one results in slightly better results in my game (which has about 50% invalid moves on average), and slightly faster training (~5%) but I didn't tried with Othello though.