When should I consider parallel search ?

Discussion of chess software programming and technical issues.

Moderators: bob, hgm, Harvey Williamson

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
syzygy
Posts: 4582
Joined: Tue Feb 28, 2012 10:56 pm

Re: When should I consider parallel search ?

Post by syzygy » Sun Jun 11, 2017 9:42 pm

elpapa wrote:Now, what aspect of solving Rubik's cube would you say I missed?
Watch carefully:
https://www.youtube.com/watch?v=R07JiT0PlcE

Pio
Posts: 157
Joined: Sat Feb 25, 2012 9:42 pm
Location: Stockholm
Contact:

Re: When should I consider parallel search ?

Post by Pio » Sun Jun 11, 2017 9:42 pm

Hej Patrik!

I would not say that you missed a single thing. I would say that you are a genius that could solve so many Rubriks' cubes at once. SMP when it is at its best :)

/Pio

PS
Jag tror att du skulle förbättra din automatiska Texel tuning om du ändrade till att minimera det absoluta felet och inte det kvadratiska. Det finns ingen logik bakom varför tio "en tiondels" så små fel skulle vara bättre än ett stort fel. Nu försöker du ta bort de riktigt stora felen vilket är de taktiska felen som en djupare sökning ändå kommer hitta

elpapa
Posts: 203
Joined: Sun Jan 18, 2009 10:27 pm
Location: Sweden
Full name: Patrik Karlsson

Re: When should I consider parallel search ?

Post by elpapa » Sun Jun 11, 2017 10:49 pm

Pio wrote:Hej Patrik!

I would not say that you missed a single thing. I would say that you are a genius that could solve so many Rubriks' cubes at once. SMP when it is at its best :)
Yes, I call it thinking outside the cube :)
Pio wrote: PS
Jag tror att du skulle förbättra din automatiska Texel tuning om du ändrade till att minimera det absoluta felet och inte det kvadratiska. Det finns ingen logik bakom varför tio "en tiondels" så små fel skulle vara bättre än ett stort fel. Nu försöker du ta bort de riktigt stora felen vilket är de taktiska felen som en djupare sökning ändå kommer hitta
Jag använder mig inte av Texel tuning och kan inte minnas att jag skrivit något inlägg om det, men tack för tipset :D

elpapa
Posts: 203
Joined: Sun Jan 18, 2009 10:27 pm
Location: Sweden
Full name: Patrik Karlsson

Re: When should I consider parallel search ?

Post by elpapa » Sun Jun 11, 2017 11:47 pm

syzygy wrote:
elpapa wrote:Now, what aspect of solving Rubik's cube would you say I missed?
Watch carefully:
https://www.youtube.com/watch?v=R07JiT0PlcE
Is this the same guy whose sweater caught fire while he was knitting it?

petero2
Posts: 605
Joined: Mon Apr 19, 2010 5:07 pm
Location: Sweden
Contact:

Re: When should I consider parallel search ?

Post by petero2 » Wed Jun 14, 2017 6:41 pm

Pio wrote:Jag tror att du skulle förbättra din automatiska Texel tuning om du ändrade till att minimera det absoluta felet och inte det kvadratiska. Det finns ingen logik bakom varför tio "en tiondels" så små fel skulle vara bättre än ett stort fel. Nu försöker du ta bort de riktigt stora felen vilket är de taktiska felen som en djupare sökning ändå kommer hitta
I have tried that but the result was very bad, around -130 elo in self play.

When optimizing a single parameter, using the L1 norm would lead to picking the parameter value that represents the median of the input data, while using the L2 norm would lead to picking the parameter value that represents the arithmetic mean of the input data. It appears a similar effect occurs when optimizing more than one parameter.

For example, I have a set of parameters that define a "drawishness" factor as a function of the half move clock. After optimizing using the L1 norm, all parameters corresponding to HMC > 40ply became zero, causing all such positions to be evaluated as exactly zero. This is expected if more than 50% of such positions are actually draws and the optimization tends to select based on the median value. For engine strength this seems really bad though.

Post Reply