I did look into this actually. It would be a bit more difficult than Crafty and Stockfish, because of the probabilistic eval. All the score margins in search (futility, razoring, etc) would have to be changed.
I am sure you considered this but what would be the issue with using the simple conversion formula
It would work for the middle linear region, but I don't know how well that will work for more extreme scores.
It may be possible to use the input instead of output of the final activation function. Then it should be more or less linear.
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.
Do you have data on the elo gain per doubling of hardware speed with Giraffe?
Normally it is somewhere around 65, but I don't know how neural nets work, so I'd be interested to know.
Presumably giving Giraffe more time only affects the search (as with a normal engine) - the eval is already fixed. Is that right?
The eval is already fixed, so during play, it should scale similarly to other engines.
I don't have the numbers, though. Sorry.
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.
I'm running Giraffe 7119bddd8913 as a Winboard engine (through Wine) in Scid. It works in infinite analysis but stops at depth 29. So I make a move (such as the one it suggests) and it starts analysing from there but stops at depth 29 again. And so on, consistently.
Marek Soszynski wrote:I'm running Giraffe 7119bddd8913 as a Winboard engine (through Wine) in Scid. It works in infinite analysis but stops at depth 29. So I make a move (such as the one it suggests) and it starts analysing from there but stops at depth 29 again. And so on, consistently.
"Depth 29" is the current hard limit for node budget due to concerns with overflow. What kind of position are you looking at? I have never seen it actually getting to "depth 29" in a non-mate position.
Also, there will be a source release soon, so you'll be able to compile a native version shortly. I do almost all my testing on Linux and OSX, and have never encountered that.
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.
Marek Soszynski wrote:I'm running Giraffe 7119bddd8913 as a Winboard engine (through Wine) in Scid. It works in infinite analysis but stops at depth 29. So I make a move (such as the one it suggests) and it starts analysing from there but stops at depth 29 again. And so on, consistently.
"Depth 29" is the current hard limit for node budget due to concerns with overflow. What kind of position are you looking at? I have never seen it actually getting to "depth 29" in a non-mate position.
Also, there will be a source release soon, so you'll be able to compile a native version shortly. I do almost all my testing on Linux and OSX, and have never encountered that.
Here's one of the positions I was studying:
[d]3r4/1pr1q1k1/p2n4/P2Pppb1/N6p/1B1R3P/5PP1/3RQ1K1 w - - 0 37
I'm certainly looking forward to a native compile of your most interesting engine.
Marek Soszynski wrote:I'm running Giraffe 7119bddd8913 as a Winboard engine (through Wine) in Scid. It works in infinite analysis but stops at depth 29. So I make a move (such as the one it suggests) and it starts analysing from there but stops at depth 29 again. And so on, consistently.
"Depth 29" is the current hard limit for node budget due to concerns with overflow. What kind of position are you looking at? I have never seen it actually getting to "depth 29" in a non-mate position.
Also, there will be a source release soon, so you'll be able to compile a native version shortly. I do almost all my testing on Linux and OSX, and have never encountered that.
Here's one of the positions I was studying:
[d]3r4/1pr1q1k1/p2n4/P2Pppb1/N6p/1B1R3P/5PP1/3RQ1K1 w - - 0 37
I'm certainly looking forward to a native compile of your most interesting engine.
Yeah that position works fine for me (on OSX). It may be something to do with Wine or Scid. Do you have an engine communication log by any chance? And can you try searching the position from command line?
post
setboard 3r4/1pr1q1k1/p2n4/P2Pppb1/N6p/1B1R3P/5PP1/3RQ1K1 w - - 0 37
analyze
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.
matthewlai wrote:
Yeah that position works fine for me (on OSX). It may be something to do with Wine or Scid. Do you have an engine communication log by any chance? And can you try searching the position from command line?
Meanwhile the behaviour in Scid, with any position, is the same. Giraffe stops analysing (as confirmed by Task Manager) when Scid shows depth 29 is reached. This could take 25 seconds or 125, to give you an idea.
matthewlai wrote:
Yeah that position works fine for me (on OSX). It may be something to do with Wine or Scid. Do you have an engine communication log by any chance? And can you try searching the position from command line?
Meanwhile the behaviour in Scid, with any position, is the same. Giraffe stops analysing (as confirmed by Task Manager) when Scid shows depth 29 is reached. This could take 25 seconds or 125, to give you an idea.
I have absolutely no idea how that can happen!
Is there a way to make Scid produce an engine communication log?
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.
