ChessAlpha Zero development

[Zeta36]

Hello, @mokemokechicken and @Yuhang.

As I promised I've done (just in one day, I had no more time) an adaptation of the reversi-zero project @mokemokechicken did into a chess version: https://github.com/Zeta36/chess-alpha-zero

The project is already functional (in the sense that it doesn't fail and the three workers do their job), but unfortunately I have no GPU (just an Intel i5 CPU) nor money to spend in a AWS server or similar.
So I just could check the self-play with the toy config "--type mini". Moreover, I had to descend self.simulation_num_per_move = 2 and self.parallel_search_num = 2.

In this way I was able to generate the 100 games needed for the optimitazion worker to start. The optimization process seemed to work perfeclty, and the model was able to reach a loss of ~0.6 after 1000 steps. I guess so that the model was able to overfit the 100 games of the former self-play.

Then I execute the evaluation process and it worked fine. The overfitted model was able to defeat the random original model of the beggining by 100% (causality??).
Finally I check the ASCII way to play against the best model. It worked as expected. To indicate our moves we have to use UCI notation: a1a2, b3b8, etc. More info here: https://chessprogramming.wikispaces.com/Algebraic+Chess+Notation

By the way, the model output is now of size 8128 (instead of the 64 of reversi and the 362 of Go), and it correspond to all possible legal UCI moves in a chess game. I generate these new labels in the config.py file.
I have to note you also that the board state (and the player turn) is traced by the fen chess notation: https://en.wikipedia.org/wiki/Forsyth%E2%80%93Edwards_Notation

Here is for example the FEN for the starting position: rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1 (w is for white to move).

I changed also a little bit the resign function. Chess is not like Go or Reversi where you always finish the game more or less in the same number of movements. In Chess the game can end in a lot of ways (checkmate, stalemate, etc.) and the self-play coul be more than 200 movements before reaching an ending position (normally in a draw). So I decided to cut-off the play after some player has more than 13 points of advantage (this score is computed as usual taking into account the value of the pieces: he queen is worth 10, roots 5.5, etc).

As you can imagine with my poor machine I could not fully test the project beyond these tiny tests of functionality. So I'd really appreciate if you could please take some free time of your GPU's for testing this implementation in a more serious way. Both you can of course be colaborators of the project if you wish.

Also I don't know if I commited some theoretical bugs after this adaptation to chess and I'd apretiate too any comments by your side in this sense.

Best regards!!

[yhyu13]

@mokemokechicken @Zeta36

Hi, guys. I refactored my APV_MCTS.py now it uses hash table the way you guys did. But I found a confusing that var_u is initialized but never used in the ReversiPlayer. Puls, my hash table will initialize a big chunk of numpy array [5x362] for N,W,Q,U,P. The code runs much faster because of the data structure advantage in comparison to the previous tree data structure. Thanks for the inspiration! It is also easy to prune the hash table if key includes the depth of game. Have you guys found memory exhaustion?

[mokemokechicken]

Hi @Zeta36

As I promised I've done (just in one day, I had no more time) an adaptation of the reversi-zero project @mokemokechicken did into a chess version: https://github.com/Zeta36/chess-alpha-zero

It's great!! Besides, making it in just one day!

Then I execute the evaluation process and it worked fine. The overfitted model was able to defeat the random original model of the beggining by 100% (causality??).

Oh, it is similar to my case.

I changed also a little bit the resign function. Chess is not like Go or Reversi where you always finish the game more or less in the same number of movements. In Chess the game can end in a lot of ways (checkmate, stalemate, etc.) and the self-play coul be more than 200 movements before reaching an ending position (normally in a draw). So I decided to cut-off the play after some player has more than 13 points of advantage (this score is computed as usual taking into account the value of the pieces: he queen is worth 10, roots 5.5, etc).

I agree your method of using another trusted judgement.
The threshold is always difficult problem...

As you can imagine with my poor machine I could not fully test the project beyond these tiny tests of functionality. So I'd really appreciate if you could please take some free time of your GPU's for testing this implementation in a more serious way. Both you can of course be colaborators of the project if you wish.
Also I don't know if I commited some theoretical bugs after this adaptation to chess and I'd apretiate too any comments by your side in this sense.

Unfortunately, I don't have free GPU time now...
However, I would like to run it if my GPU has free time.

FYI:
Also, I applied to https://www.tensorflow.org/tfrc/. I do not know if it will be accepted. . .

[mokemokechicken]

Hi @yhyu13

Thanks for the message!
However, I would appreciate it if you divide ISSUE by topic.

Hi, guys. I refactored my APV_MCTS.py now it uses hash table the way you guys did. But I found a confusing that var_u is initialized but never used in the ReversiPlayer.

Yes, that's right. It is unnecessary.
I thought about refactoring but I forgot...

Puls, my hash table will initialize a big chunk of numpy array [5x362] for N,W,Q,U,P. The code runs much faster because of the data structure advantage in comparison to the previous tree data structure. Thanks for the inspiration!

Oh, it is wonderful!

Have you guys found memory exhaustion?

No, I haven't.
The amount of memory consumed is proportional to the number of expanded nodes.
In my reversi, the number of explorations is small (~ 500), and since the number of moves is small, it was hardly a problem (Up to about 500 * 64).

[Zeta36]

Hello, @mokemokechicken and @yhyu13 .

I've done today a new version of the Reversi Zero project. This time I adapted it to the game Connect4: https://github.com/Zeta36/connect4-alpha-zero

@mokemokechicken I'm really in love with your implementation. You did it (and DeepMind thought it) in a way that I can apply it easily to any new environment I imagine.

Moreover, Connect4 is a more easy game and I could train the model without GPU. Results are amazing. The model is able to learn to play well in only 3 generations in a couple of hours (just with a Intel i5 CPU).

I insist you've done a GREAT work, friend.

It's a pitty I don't have enough power machine to check if the chess version is able to learn to play well.

[gooooloo]

@Zeta36 good job!

Just one thinking: maybe Connect4 is easy enough so it even doesn't need an accurate NN? Just MCST will make AI strong enough?

Besides, Connect4 is a very unbalanced game. Who plays first should always win, which makes AI training much easier. So how about Connect5 with specific opening pattern, such as "11D (瑞星): equal" in https://en.wikipedia.org/wiki/Renju_opening_pattern. That seems to be an more persuasive game for AlphaGoZero algo.

[mokemokechicken]

@Zeta36

Great!

I was thinking about trying it with that game too!
However, I knew for the first time the name Connect4 (^^;
I am happy to see your implementation of various games!!

Moreover, Connect4 is a more easy game and I could train the model without GPU. Results are amazing. The model is able to learn to play well in only 3 generations in a couple of hours (just with a Intel i5 CPU).

That's interesting and a nice property.

Just thinking:
As @gooooloo and @yhyu13 said,
I am interested in "the version using only MCTS" and "the version using only policy move(simulation_num_per_move=1)".