Hi.
During my development of Loki's neural network evaluation attempt, I stumbled over the Neural MoveMap Heuristic on CPW. The idea seems really interesting, but I am wondering if anyone else has tried it? If so, how did it go?
Neural MoveMap
Moderators: hgm, Rebel, chrisw
-
- Posts: 174
- Joined: Thu Nov 26, 2020 10:06 am
- Full name: Niels Abildskov
-
- Posts: 13
- Joined: Fri Dec 18, 2020 7:58 pm
- Full name: Lech Mazur
Re: Neural MoveMap
You can tell it's a 20-year-old paper when a neural net with 40 hidden neurons is described as "very large."
-
- Posts: 174
- Joined: Thu Nov 26, 2020 10:06 am
- Full name: Niels Abildskov
Re: Neural MoveMap
Well yeah, but old doesn't necessarily mean bad... NMP is also an old method
-
- Posts: 438
- Joined: Mon Apr 24, 2006 8:06 pm
Re: Neural MoveMap
This idea could be applied to a modern convolutional neural network as well. It is a way to encode the policy output.
Whether it would perform better or not than the AlphaZero policy encoding is difficult to predict. I guess it would not make a big difference, especially if the network is big.
Whether it would perform better or not than the AlphaZero policy encoding is difficult to predict. I guess it would not make a big difference, especially if the network is big.
-
- Posts: 16
- Joined: Fri Dec 27, 2019 8:47 pm
- Full name: Jacek Dermont
Re: Neural MoveMap
There is (was?) some discussion on stockfish discord about using separate neural network or additional head for move ordering, but it was more about likelihood of cutoff. Given that we already have relatively fast and somewhat reliable value network, it is interesting to see if relatively fast pseudo-policy network for move ordering would yield even better results.
-
- Posts: 174
- Joined: Thu Nov 26, 2020 10:06 am
- Full name: Niels Abildskov
Re: Neural MoveMap
That makes sense.derjack wrote: ↑Wed May 19, 2021 12:58 pm There is (was?) some discussion on stockfish discord about using separate neural network or additional head for move ordering, but it was more about likelihood of cutoff. Given that we already have relatively fast and somewhat reliable value network, it is interesting to see if relatively fast pseudo-policy network for move ordering would yield even better results.
If I understand the cpw page correctly, the original MoveMap was used to give somewhat of a “meta-history” in the sense that it would be used before searching to get good history scores early on, which would help move ordering.
This is probably the same as a probability of cutoff since bigger scores would mean the move was likely to be better.