Are kiudee parameters the best?

Discussion of anything and everything relating to chess playing software and machines.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
corres
Posts: 2907
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Are kiudee parameters the best?

Post by corres » Thu Jan 30, 2020 2:00 pm

Hai wrote:
Thu Jan 30, 2020 1:05 pm
M ANSARI wrote:
Thu Jan 30, 2020 7:41 am
Wouldn't it be possible to have an automated system check that looks at your computer hardware and then follow a set table to automatically tune your Lc0 to the best possible settings? I mean it should be easy enough to poll the computer for number of available cores and frequency and then also check the memory and GPU. This could be a tuning app that runs once you install the Lc0 engine. It might not be a perfect tune, but for sure it would be better than the default.
Yes it is possible.
Everybody would get a 50 elo stronger LC0 now.
Maybe even 100 elo stronger when using 2 gpus.
You are too optimistic.
Tests prove only in special situations kiudee setting gives important chess power enhancement.
In general situations kiudee may causes even weakening in chess power.

Hai
Posts: 548
Joined: Sun Aug 04, 2013 11:19 am

Re: Are kiudee parameters the best?

Post by Hai » Thu Jan 30, 2020 2:47 pm

corres wrote:
Thu Jan 30, 2020 2:00 pm
Hai wrote:
Thu Jan 30, 2020 1:05 pm
M ANSARI wrote:
Thu Jan 30, 2020 7:41 am
Wouldn't it be possible to have an automated system check that looks at your computer hardware and then follow a set table to automatically tune your Lc0 to the best possible settings? I mean it should be easy enough to poll the computer for number of available cores and frequency and then also check the memory and GPU. This could be a tuning app that runs once you install the Lc0 engine. It might not be a perfect tune, but for sure it would be better than the default.
Yes it is possible.
Everybody would get a 50 elo stronger LC0 now.
Maybe even 100 elo stronger when using 2 gpus.
You are too optimistic.
Tests prove only in special situations kiudee setting gives important chess power enhancement.
In general situations kiudee may causes even weakening in chess power.
We weren't talking about kiudee settings.

corres
Posts: 2907
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Are kiudee parameters the best?

Post by corres » Thu Jan 30, 2020 3:07 pm

Hai wrote:
Thu Jan 30, 2020 2:47 pm
corres wrote:
Thu Jan 30, 2020 2:00 pm
Hai wrote:
Thu Jan 30, 2020 1:05 pm
M ANSARI wrote:
Thu Jan 30, 2020 7:41 am
Wouldn't it be possible to have an automated system check that looks at your computer hardware and then follow a set table to automatically tune your Lc0 to the best possible settings? I mean it should be easy enough to poll the computer for number of available cores and frequency and then also check the memory and GPU. This could be a tuning app that runs once you install the Lc0 engine. It might not be a perfect tune, but for sure it would be better than the default.
Yes it is possible.
Everybody would get a 50 elo stronger LC0 now.
Maybe even 100 elo stronger when using 2 gpus.
You are too optimistic.
Tests prove only in special situations kiudee setting gives important chess power enhancement.
In general situations kiudee may causes even weakening in chess power.
We weren't talking about kiudee settings.
From parameters yielded from automated tuning you can not hope much more than kiudee settings.

User avatar
M ANSARI
Posts: 3435
Joined: Thu Mar 16, 2006 6:10 pm

Re: Are kiudee parameters the best?

Post by M ANSARI » Thu Jan 30, 2020 3:37 pm

Hai wrote:
Thu Jan 30, 2020 2:47 pm
corres wrote:
Thu Jan 30, 2020 2:00 pm
Hai wrote:
Thu Jan 30, 2020 1:05 pm
M ANSARI wrote:
Thu Jan 30, 2020 7:41 am
Wouldn't it be possible to have an automated system check that looks at your computer hardware and then follow a set table to automatically tune your Lc0 to the best possible settings? I mean it should be easy enough to poll the computer for number of available cores and frequency and then also check the memory and GPU. This could be a tuning app that runs once you install the Lc0 engine. It might not be a perfect tune, but for sure it would be better than the default.
Yes it is possible.
Everybody would get a 50 elo stronger LC0 now.
Maybe even 100 elo stronger when using 2 gpus.
You are too optimistic.
Tests prove only in special situations kiudee setting gives important chess power enhancement.
In general situations kiudee may causes even weakening in chess power.
We weren't talking about kiudee settings.

Yes of course I mean this is for improving on the default setting. Lc0 seems very different in that depending on the GPU or GPU's it uses and the possible NPS it can get, then the start parameters need to be adjusted for best results. With AB engines this seems to not be the case or at least the default parameters are usually pretty close to optimal (except maybe Time Management for different TC's).

mwyoung
Posts: 2080
Joined: Wed May 12, 2010 8:00 pm

Re: Are kiudee parameters the best?

Post by mwyoung » Thu Jan 30, 2020 7:16 pm

M ANSARI wrote:
Thu Jan 30, 2020 3:37 pm
Hai wrote:
Thu Jan 30, 2020 2:47 pm
corres wrote:
Thu Jan 30, 2020 2:00 pm
Hai wrote:
Thu Jan 30, 2020 1:05 pm
M ANSARI wrote:
Thu Jan 30, 2020 7:41 am
Wouldn't it be possible to have an automated system check that looks at your computer hardware and then follow a set table to automatically tune your Lc0 to the best possible settings? I mean it should be easy enough to poll the computer for number of available cores and frequency and then also check the memory and GPU. This could be a tuning app that runs once you install the Lc0 engine. It might not be a perfect tune, but for sure it would be better than the default.
Yes it is possible.
Everybody would get a 50 elo stronger LC0 now.
Maybe even 100 elo stronger when using 2 gpus.
You are too optimistic.
Tests prove only in special situations kiudee setting gives important chess power enhancement.
In general situations kiudee may causes even weakening in chess power.
We weren't talking about kiudee settings.

Yes, of course I mean this is for improving on the default setting. Lc0 seems very different in that depending on the GPU or GPU's it uses and the possible NPS it can get, then the start parameters need to be adjusted for best results. With AB engines this seems to not be the case or at least the default parameters are usually pretty close to optimal (except maybe Time Management for different TC's).
Yes, NN do not use a A/B search. It is a totally different animal. And until you can rap your head around how the NN engines search the game tree. It can be a bit of a mystery, but only because it is new. Maybe we can clarify a bit....

The focus of MCTS is on the analysis of the most promising moves, expanding the search tree based on random sampling of the search space. The application of Monte Carlo tree search in games is based on many playouts also called roll-outs. In each playout, the game is played out to the very end by selecting moves at random. The final game result of each playout is then used to weight the nodes in the game tree so that better nodes are more likely to be chosen in future playouts.

The A/B search is more intuitive...
Professing themselves to be wise, they became fools,
take on me. Foes 0.

User avatar
M ANSARI
Posts: 3435
Joined: Thu Mar 16, 2006 6:10 pm

Re: Are kiudee parameters the best?

Post by M ANSARI » Sat Feb 01, 2020 7:43 am

[/quote]
The A/B search is more intuitive...
[/quote]


That is why I still think the perfect solution is an engine that combines A/B with NN. I know that nobody yet is looking at that, but it might make for an interesting fork for SF if implemented correctly. Of course I am following the impressive progress of Lc0, but it still plays poor moves in some positions that are way below its ELO strength. The thing is that NN engines are basically GPU engines and for most computers that means that the CPU cores available are not being utilized fully. That is like a car that is running on half it cylinders ... if somehow in the background a strong A/B engine is running and it sees a fail high on a move or a big jump in a move ... that move would be thrown into the NN engine to check and play out the line. Usually when Lc0 plays a bad move it will have a fail high a few moves after ... or a good move is seen as a good move a few moves later. So if the engine can play out that line against its own A/B engine it can then decide if the move is better than its original move it wanted to play.

jp
Posts: 1345
Joined: Mon Apr 23, 2018 5:54 am

Re: Are kiudee parameters the best?

Post by jp » Sat Feb 01, 2020 8:47 am

mwyoung wrote:
Thu Jan 30, 2020 7:16 pm
In each playout, the game is played out to the very end by selecting moves at random.
This is not correct, right? In game-play (not training), most playouts will not end in terminal nodes.

corres
Posts: 2907
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Are kiudee parameters the best?

Post by corres » Sat Feb 01, 2020 9:03 am

M ANSARI wrote:
Sat Feb 01, 2020 7:43 am

That is why I still think the perfect solution is an engine that combines A/B with NN. I know that nobody yet is looking at that, but it might make for an interesting fork for SF if implemented correctly. Of course I am following the impressive progress of Lc0, but it still plays poor moves in some positions that are way below its ELO strength. The thing is that NN engines are basically GPU engines and for most computers that means that the CPU cores available are not being utilized fully. That is like a car that is running on half it cylinders ... if somehow in the background a strong A/B engine is running and it sees a fail high on a move or a big jump in a move ... that move would be thrown into the NN engine to check and play out the line. Usually when Lc0 plays a bad move it will have a fail high a few moves after ... or a good move is seen as a good move a few moves later. So if the engine can play out that line against its own A/B engine it can then decide if the move is better than its original move it wanted to play.
The main issue is not only the NN engines but AB engines also make bad moves so we would have a super-sagacious entity to make decision about what move would play: the move of NN engines or the move of AB engines. But in the most cases even a super-GM can not supervise the move of engines.

User avatar
M ANSARI
Posts: 3435
Joined: Thu Mar 16, 2006 6:10 pm

Re: Are kiudee parameters the best?

Post by M ANSARI » Sat Feb 01, 2020 9:51 am

I think you misunderstood ... the idea would be that the NN engine would still chose the move ... the A/B engine move would only be a suggestion for the NN engine to use as one of its moves it should search. So in the end if that move the NN does not agree is better (after playing out the line in search) then the NN will still keep its original move it wanted to play. The cost of searching this extra line for the NN engine would be very minimal in GPU cycles as the depth would be relatively shallow (probably less than 10 ply). The gain would be that tactics that are missed by NN due to not having it in its learned games are covered. So theoretically the NN engine would still play the same but the only thing that would change is that it might have an extra few lines to search (suggested by the A/B engine) that it would have otherwise not searched.

corres
Posts: 2907
Joined: Wed Nov 18, 2015 10:41 am
Location: hungary

Re: Are kiudee parameters the best?

Post by corres » Sat Feb 01, 2020 11:55 am

M ANSARI wrote:
Sat Feb 01, 2020 9:51 am
I think you misunderstood ... the idea would be that the NN engine would still chose the move ... the A/B engine move would only be a suggestion for the NN engine to use as one of its moves it should search. So in the end if that move the NN does not agree is better (after playing out the line in search) then the NN will still keep its original move it wanted to play. The cost of searching this extra line for the NN engine would be very minimal in GPU cycles as the depth would be relatively shallow (probably less than 10 ply). The gain would be that tactics that are missed by NN due to not having it in its learned games are covered. So theoretically the NN engine would still play the same but the only thing that would change is that it might have an extra few lines to search (suggested by the A/B engine) that it would have otherwise not searched.
So you think NN engines can work as an "supervisor".
I think no NN engines and no AB engines either can work as a "superwisor" because both of them have the own weak points.
Moreover what you wrote down is the base of LeelaFish. LeelaFish is give some plus Elo if the used GPU is weak (GTX-types).

Post Reply