AlphaZero

[Fafkorn]

Hello, Im trying to develop my own Alpha Zero engine. In the original paper authors are using board representation with information about previous moves. My goal is to achieve engine about 2000 ELO, I don't have too much expectations. Using standard 773 bits board representation (12x8x8 + castling + move) would work? Do I need pretrain that model? Do I have to use convolutional layers? Thanks in advance.

[brianr]

Suggest reviewing this first:

https://github.com/Zeta36/chess-alpha-zero

or here for a simpler game like Connect4:

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

[Fafkorn]

I'm on the stage where I implemented MCTS and I have some random NN with few layers just to test working.
Zeta36 is using input of size (18, 8, 8)
18 - 12 piece type, 4 castling rights, 1 for fifty moves rule, 1 for en-passant

Information about side to move isn't necessary?
Author is using many many layers, doesn't it affect on the performance?
So many bits used for castling, en passant and fifty moves rules are used only because of normalization, right?

[brianr]

Sometimes the answer is simply because that's what Alpha Zero did...

Experiment and see what works for you after you have established a working baseline.

Have fun.

[Fafkorn]

Pretraining that author executed is based on grandmaster games. Policy output was hot-one for move chosen by GM?

[supersharp77]

Pretraining that author executed is based on grandmaster games. Policy output was hot-one for move chosen by GM?

NN Policy Output....



https://towardsdatascience.com/policy-n ... 2776056ad2

http://web.mst.edu/~gosavia/neural_networks_RL.pdf

https://www.researchgate.net/publicatio ... y_Gradient

https://people.eecs.berkeley.edu/~svlev ... mfcgps.pdf

https://flyyufelix.github.io/2017/10/12/dqn-vs-pg.html

All the top GM (and master games) should be integrated into the learning process..and if successful engine strength should be at a minimum 2500+ (See Giraffe chess engine)....research models are a bit complex....good luck! AR

[smcracraft]

Skip it.

Stand up lczero.org and enjoy.

No need to reinvent the wheel.

Leela later will be more Tal-like once they fix it.

Stuart

[phhnguyen]

Skip it.

Stand up lczero.org and enjoy.

No need to reinvent the wheel.

Do you want to divide the computer chess world into only two groups: lc0 clone and traditional alpha-beta?

[Fafkorn]

Skip it.

Stand up lczero.org and enjoy.

No need to reinvent the wheel.

Leela later will be more Tal-like once they fix it.

Stuart

I'm not trying to compete with LeelaChessZero or AlphaZero. I'm trying to develop my own ALphaZero for academic purpose (my Thesis). I just want to dispel some doubts.

[Fafkorn]

I have another question according to this topic. When I get my policy from neural network (in my case 4096 numbers), let's say only 3 moves are legal with policy values 0.1, 0.1, 0.1. This affects on proportions postion reward to visits value in my U(s, a). Do I have to normalize policy values and how can I do this?