Houdini 3 reducing the depth feature

[Rebel]

He does not have a 10 node cluster.

Just for the record, I don't need that to test a 20-30 elo difference, a couple of thousands games will do just fine to proof the PVS extension better at nowadays hardware. It's what I did last year and it (still) showed a remarkable elo difference, enough to abort the matches.

[Karlo Bala]

Matthew, I am not sure you have understood the danger you are heading to full stream These 2 guys were already enjoying fighting each other well before you were borne.

Gold rule for new members: read the old CCC archives first
http://www.stmintz.com/ccc/

And the root of evil: http://www.rebel.nl/match.htm

[bob]

Bob,

Ed is a 70-odd year old man.

He is retired.

He does not have the hyperfast testing facilities that you have.

He does not have a 10 node cluster.

He does not have a crappy IBRIX filesystem.

He does not have a Cray supercomputer lying around, etc. etc.

He is not the richest person in the world.

He is not able to afford your facilities.

So, unless you want to test every last change since REBEL 1, let him be.

Matthew:out

1. IBRIX is not a "crappy filesystem". However, we don't have IBRIX either, so I don't quite see what your point is here...

2. I don't have a Cray lying about either. Again, what is your point?

My point was that for the longest, as reported by Ed in the 90's, his testing approach was to play maybe 8 games at a time (LONG games) and use maybe 40 games to decide whether a change was good or bad. That is random noise.

I don't believe this extension is worth 30 Elo, because the check extension is not worth 30 Elo today. If one makes a claim, one should be able to back it up with actual testing results, otherwise it is nothing more than "idle speculation".

Butt out...

[Uri Blass]

Bob,

Ed is a 70-odd year old man.

He is retired.

He does not have the hyperfast testing facilities that you have.

He does not have a 10 node cluster.

He does not have a crappy IBRIX filesystem.

He does not have a Cray supercomputer lying around, etc. etc.

He is not the richest person in the world.

He is not able to afford your facilities.

So, unless you want to test every last change since REBEL 1, let him be.

Matthew:out

1. IBRIX is not a "crappy filesystem". However, we don't have IBRIX either, so I don't quite see what your point is here...

2. I don't have a Cray lying about either. Again, what is your point?

My point was that for the longest, as reported by Ed in the 90's, his testing approach was to play maybe 8 games at a time (LONG games) and use maybe 40 games to decide whether a change was good or bad. That is random noise.

I don't believe this extension is worth 30 Elo, because the check extension is not worth 30 Elo today. If one makes a claim, one should be able to back it up with actual testing results, otherwise it is nothing more than "idle speculation".

Butt out...

I think that the value of extensions is dependent on the program.
There are different programs than Crafty and I do not believe that the check extension is not worth 30 elo today for everybody and there may be people that it worth more than 30 elo for them.

Ed said that he found 20-30 elo advantage today for the PVS extension based on some thousands of games.

[Rebel]

Bob,

Ed is a 70-odd year old man.

He is retired.

He does not have the hyperfast testing facilities that you have.

He does not have a 10 node cluster.

He does not have a crappy IBRIX filesystem.

He does not have a Cray supercomputer lying around, etc. etc.

He is not the richest person in the world.

He is not able to afford your facilities.

So, unless you want to test every last change since REBEL 1, let him be.

Matthew:out

1. IBRIX is not a "crappy filesystem". However, we don't have IBRIX either, so I don't quite see what your point is here...

2. I don't have a Cray lying about either. Again, what is your point?

It was a post in support of you

My point was that for the longest, as reported by Ed in the 90's, his testing approach was to play maybe 8 games at a time (LONG games) and use maybe 40 games to decide whether a change was good or bad. That is random noise.

Wrong again. Now if you show some kind of minimum respect for your peers (me this time) I might post some stuff how testing with limited hardware (by myself and others) was done back then.

I don't believe this extension is worth 30 Elo, because the check extension is not worth 30 Elo today. If one makes a claim, one should be able to back it up with actual testing results, otherwise it is nothing more than "idle speculation".

Patronizing someone else work usually works counter productive, if people give you the middle finger don't be too surprised. I owe you nothing.

As for the PVS extension, it's the result of months of work finding the right formula and parameters, it's explained on my page open for every engine programmer to try if there is some elo gain in it for them.

[Cardoso]

This approach makes more sense to me for the fail-low situations, at least from the point of the ELO increase.
Did you try it for fail-lows?

I guess right after your idea everyone made a run on testing the fail low case
It also occurred to me the same idea. Unfortunately my personal life didn't allow me to test even the FH case.
But I would assume on the FL case it can help even if it must be done slightly different.

best regards,
Alvaro

[jefk]

[quote="Cardoso"][quote="Dragan"]This approach makes more sense to me for the fail-low situations, at least from the point of the ELO increase.
Did you try it for fail-lows?[/quote]
I guess right after your idea everyone made a run on testing the fail low case
Alvaro[/quote]

well i didn't read this whole discussion in detail, but as engine user in practical games obviously the time factor usually is important, so when eg. at higher depths your time per move (eg three minutes for long game) is getting too long, some engines 'force' a move, and then i guess its a useful feature to have a slightly better estimate of the best move available, even although it might not be perfect when completing the whole search at the additional one more ply.
just my 2 cnts