A group of researchers have trained recurrent neural networks to learn how to perform a number of algorithms, including Monte-Carlo tree search for chess which is used in a number of engines such as Dragon, Leela, and Shashchess (Stockfish derivative), without handcrafting the search algorithm itself.
https://arxiv.org/abs/2106.04537
Learning search algorithms using neural networks
Moderators: hgm, Rebel, chrisw
-
- Posts: 512
- Joined: Tue Sep 29, 2020 4:29 pm
- Location: Dublin, Ireland
- Full name: Madeleine Birchfield
-
- Posts: 4190
- Joined: Wed Nov 25, 2009 1:47 am
Re: Learning search algorithms using neural networks
Totally useless work. The typical example of goal of some of academic community being publishing for the sake of publishing and nothing else.Madeleine Birchfield wrote: ↑Sat Jul 10, 2021 7:14 pm A group of researchers have trained recurrent neural networks to learn how to perform a number of algorithms, including Monte-Carlo tree search for chess which is used in a number of engines such as Dragon, Leela, and Shashchess (Stockfish derivative), without handcrafting the search algorithm itself.
https://arxiv.org/abs/2106.04537
-
- Posts: 358
- Joined: Thu Jan 22, 2015 3:21 pm
- Location: Zurich, Switzerland
- Full name: Jonathan Rosenthal
Re: Learning search algorithms using neural networks
The paper looks interesting, but by my understanding they are not doing anything nearly as complex as learning to perform Monte Carlo Tree search. Instead they are merely showing that you can train an RNN to successfully scale over more iterations. I am neither particularly surprised nor impressed by the results, but to me it is still an interesting read. The paper also seems unfinished as they have a section for more difficult chess problems that is not yet written.Madeleine Birchfield wrote: ↑Sat Jul 10, 2021 7:14 pm A group of researchers have trained recurrent neural networks to learn how to perform a number of algorithms, including Monte-Carlo tree search for chess which is used in a number of engines such as Dragon, Leela, and Shashchess (Stockfish derivative), without handcrafting the search algorithm itself.
https://arxiv.org/abs/2106.04537
Algorithms which scale without explicit tree search are interesting to me because they could be better at things like fortresses in chess and may be optimal for domains where explicit tree search is not possible, such as fog of war chess. RNNs are the most obvious model to try for this problem, but my intuition is that it should be possible to come up with a lot better.
-Jonathan