4 ms·

Stockfish 16.1

33 points by jonbaer 3y ago

Bostonian 3y ago
'This release marks the removal of the traditional handcrafted evaluation and the transition to a fully neural network-based approach.'
- karmakaze 3y ago
  Like AlphaGo Zero, but for Chess.
  - janalsncm 3y ago
    Stockfish still runs a full alpha beta tree search, so it’s not really the same.
TylerLives 3y ago
Can Stockfish (with small modifications) be used for other games now? There are a few decent open source AlphaZero implementations and I wonder how it would compare.
- janalsncm 3y ago
  Depends on what games and what you define as “small” modifications. There are a lot of game-specific tweaks in the search code. You can still apply alpha-beta tree search to lots of games though.
  - TylerLives 3y ago
    NNUE is the interesting part to me. Alpha-beta tree search is useless without a good value function. Not sure what would be the best way to generate the training data if you're starting from scratch.
    - janalsncm 3y ago
      One way is through self-play. You can pit two value functions against each other to determine which is better, thereby learning the value function.
artninja1988 3y ago
Now that it uses NN only and does away with search, does it use more or less computing resources? Also, does it suffer from the "Swiss cheese" problem like the ones for go do? People could essentially look for weaknesses in go engines by finding paths that the engine hasn't explored during self play and the accuracy would plummet in a way that humans could beat it, as far as I understand.
- osti 3y ago
  It doesn't do away with search..
janalsncm 3y ago
I’m curious about the comparison between NNUE and the LLM-based model that Deep Mind announced a couple of weeks ago (https://arxiv.org/pdf/2402.04494.pdf https://arxiv.org/pdf/2402.04494.pdf). Using NNUE only (i.e. depth 1 search) would be directly comparable. If Deep Mind’s model is better it raises interesting questions about scaling laws for this kind of thing.
faeriechangling 3y ago
It's crazy to me how it keeps getting better when stockfish was already destroying the best players in the world with version 1.
sireat 3y ago
Complete Removal of handcrafted evaluation (HCE) is a pretty big milestone. Original strength of Stockfish was closer to Type B strategy (per Shannon) as opposed to Type A(brute force) of Deep Blue. That is Stockfish was evaluating relatively few positions / second compared to brute forcers (like Crafty,Fritz etc). This was offset by the best eval (basically crowdsourced human GM/IM/FM knowledge) heuretics. As an FM I could "exploit" Fritzes and Crafties from 1995-2005 by using holes in their eval. Tim Crabbe provides some examples from that era: https://timkr.home.xs4all.nl/chess2/honor.htm https://timkr.home.xs4all.nl/chess2/honor.htm With Stockfish, its eval was always top notch (compared to GM) and constantly improving. Obviously Stockfish was always a few orders faster than a human.

33 points by jonbaer 3y ago

'This release marks the removal of the traditional handcrafted evaluation and the transition to a fully neural network-based approach.'

karmakaze 3y ago
Like AlphaGo Zero, but for Chess.
- janalsncm 3y ago
  Stockfish still runs a full alpha beta tree search, so it’s not really the same.

TylerLives 3y ago

Can Stockfish (with small modifications) be used for other games now? There are a few decent open source AlphaZero implementations and I wonder how it would compare.

janalsncm 3y ago
Depends on what games and what you define as “small” modifications. There are a lot of game-specific tweaks in the search code. You can still apply alpha-beta tree search to lots of games though.
- TylerLives 3y ago
  NNUE is the interesting part to me. Alpha-beta tree search is useless without a good value function. Not sure what would be the best way to generate the training data if you're starting from scratch.
  - janalsncm 3y ago
    One way is through self-play. You can pit two value functions against each other to determine which is better, thereby learning the value function.

artninja1988 3y ago

Now that it uses NN only and does away with search, does it use more or less computing resources? Also, does it suffer from the "Swiss cheese" problem like the ones for go do? People could essentially look for weaknesses in go engines by finding paths that the engine hasn't explored during self play and the accuracy would plummet in a way that humans could beat it, as far as I understand.

osti 3y ago
It doesn't do away with search..

janalsncm 3y ago

I’m curious about the comparison between NNUE and the LLM-based model that Deep Mind announced a couple of weeks ago (https://arxiv.org/pdf/2402.04494.pdf https://arxiv.org/pdf/2402.04494.pdf). Using NNUE only (i.e. depth 1 search) would be directly comparable. If Deep Mind’s model is better it raises interesting questions about scaling laws for this kind of thing.

faeriechangling 3y ago

It's crazy to me how it keeps getting better when stockfish was already destroying the best players in the world with version 1.

sireat 3y ago

Complete Removal of handcrafted evaluation (HCE) is a pretty big milestone. Original strength of Stockfish was closer to Type B strategy (per Shannon) as opposed to Type A(brute force) of Deep Blue. That is Stockfish was evaluating relatively few positions / second compared to brute forcers (like Crafty,Fritz etc). This was offset by the best eval (basically crowdsourced human GM/IM/FM knowledge) heuretics. As an FM I could "exploit" Fritzes and Crafties from 1995-2005 by using holes in their eval. Tim Crabbe provides some examples from that era: https://timkr.home.xs4all.nl/chess2/honor.htm https://timkr.home.xs4all.nl/chess2/honor.htm With Stockfish, its eval was always top notch (compared to GM) and constantly improving. Obviously Stockfish was always a few orders faster than a human.