A question that I've been thinking about for a while:
Neuronets alone suffer from tactical problems because these networks are designed to answer minimum changes in the input with minimum changes in the output. Chess is a game where displacing one piece by one square can make a decisive difference, so that doesn't always blend well.
But what about an engine that does a conventional tree search with a minimum eval just to filter all tactically sound moves and feeds them to a neuronet for the final selection?
Are there engines that implement this approach? Or is the whole idea nonsense?
Neuronet plus conventional approach combined?
Moderators: hgm, Rebel, chrisw
-
- Posts: 2495
- Joined: Tue Aug 30, 2016 8:19 pm
- Full name: Rasmus Althoff
-
- Posts: 793
- Joined: Sun Aug 03, 2014 4:48 am
- Location: London, UK
Re: Neuronet plus conventional approach combined?
Giraffe uses a conventional search with a neural net evaluation function. Using a fast search to sort out tactics is another approach that I have spent quite some time thinking about, but haven't had a chance to implement.Ras wrote:A question that I've been thinking about for a while:
Neuronets alone suffer from tactical problems because these networks are designed to answer minimum changes in the input with minimum changes in the output. Chess is a game where displacing one piece by one square can make a decisive difference, so that doesn't always blend well.
But what about an engine that does a conventional tree search with a minimum eval just to filter all tactically sound moves and feeds them to a neuronet for the final selection?
Are there engines that implement this approach? Or is the whole idea nonsense?
Disclosure: I work for DeepMind on the AlphaZero project, but everything I say here is personal opinion and does not reflect the views of DeepMind / Alphabet.
-
- Posts: 2495
- Joined: Tue Aug 30, 2016 8:19 pm
- Full name: Rasmus Althoff
Re: Neuronet plus conventional approach combined?
Thanks, that's an answer from the chess-neuro-expert No. 1.
I mean, computers are better than humans with their current tactics approach, but strategically, human masters are still better. Only that humans fail tactically before their strategies might come to fruitition.
What is your take, are neuronets possibly suited better to the diffuse strategic long-term aspects of the game? Could they implement this in case they don't have to worry about tactical shots?
I mean, computers are better than humans with their current tactics approach, but strategically, human masters are still better. Only that humans fail tactically before their strategies might come to fruitition.
What is your take, are neuronets possibly suited better to the diffuse strategic long-term aspects of the game? Could they implement this in case they don't have to worry about tactical shots?
-
- Posts: 94
- Joined: Thu Feb 27, 2014 8:19 pm
Re: Neuronet plus conventional approach combined?
Spawkfish is an engine that tries to predict moves only with neural nets, and does reasonably well at that. If it could be made fast enough, using it for move ordering and heavy LMR would probably make a quite strong engine.
-
- Posts: 793
- Joined: Sun Aug 03, 2014 4:48 am
- Location: London, UK
Re: Neuronet plus conventional approach combined?
It's an interesting idea. Neural nets are certainly better at positional stuff than what we have at the moment, but they are a bit too slow to compete tactically.Ras wrote:Thanks, that's an answer from the chess-neuro-expert No. 1.
I mean, computers are better than humans with their current tactics approach, but strategically, human masters are still better. Only that humans fail tactically before their strategies might come to fruitition.
What is your take, are neuronets possibly suited better to the diffuse strategic long-term aspects of the game? Could they implement this in case they don't have to worry about tactical shots?
There are some practical issues with using 2 evaluation functions in the same search tree, so that will still require some thinking. I have some ideas on how to do it, but nothing I am really confident in yet.
Disclosure: I work for DeepMind on the AlphaZero project, but everything I say here is personal opinion and does not reflect the views of DeepMind / Alphabet.
-
- Posts: 793
- Joined: Sun Aug 03, 2014 4:48 am
- Location: London, UK
Re: Neuronet plus conventional approach combined?
Giraffe already does that. It's called the "move evaluator" net in Giraffe. It didn't help as much as I had hoped, but it did help a bit.thomasahle wrote:Spawkfish is an engine that tries to predict moves only with neural nets, and does reasonably well at that. If it could be made fast enough, using it for move ordering and heavy LMR would probably make a quite strong engine.
Disclosure: I work for DeepMind on the AlphaZero project, but everything I say here is personal opinion and does not reflect the views of DeepMind / Alphabet.
-
- Posts: 94
- Joined: Thu Feb 27, 2014 8:19 pm
Re: Neuronet plus conventional approach combined?
Cool! How strong would you estimate the move evaluator is on its own, with no search?matthewlai wrote:Giraffe already does that. It's called the "move evaluator" net in Giraffe. It didn't help as much as I had hoped, but it did help a bit.thomasahle wrote:Spawkfish is an engine that tries to predict moves only with neural nets, and does reasonably well at that. If it could be made fast enough, using it for move ordering and heavy LMR would probably make a quite strong engine.
-
- Posts: 793
- Joined: Sun Aug 03, 2014 4:48 am
- Location: London, UK
Re: Neuronet plus conventional approach combined?
Depends on what features you give it. If you give it SEE scores maybe 2000 level. If you don't, it will probably make many tactical mistakes.thomasahle wrote:Cool! How strong would you estimate the move evaluator is on its own, with no search?matthewlai wrote:Giraffe already does that. It's called the "move evaluator" net in Giraffe. It didn't help as much as I had hoped, but it did help a bit.thomasahle wrote:Spawkfish is an engine that tries to predict moves only with neural nets, and does reasonably well at that. If it could be made fast enough, using it for move ordering and heavy LMR would probably make a quite strong engine.
Disclosure: I work for DeepMind on the AlphaZero project, but everything I say here is personal opinion and does not reflect the views of DeepMind / Alphabet.
-
- Posts: 94
- Joined: Thu Feb 27, 2014 8:19 pm
Re: Neuronet plus conventional approach combined?
Sounds like something that would be fun to play against. If I run giraffe and tell it to search depth 0, will that simulate the policy network reasonably well?matthewlai wrote:Depends on what features you give it. If you give it SEE scores maybe 2000 level. If you don't, it will probably make many tactical mistakes.
-
- Posts: 793
- Joined: Sun Aug 03, 2014 4:48 am
- Location: London, UK
Re: Neuronet plus conventional approach combined?
That wouldn't work unfortunately. It would still do a QSearch. The move evaluator network is also only trained on quiet moves.thomasahle wrote:Sounds like something that would be fun to play against. If I run giraffe and tell it to search depth 0, will that simulate the policy network reasonably well?matthewlai wrote:Depends on what features you give it. If you give it SEE scores maybe 2000 level. If you don't, it will probably make many tactical mistakes.
Disclosure: I work for DeepMind on the AlphaZero project, but everything I say here is personal opinion and does not reflect the views of DeepMind / Alphabet.