TalkChess.com

Vinvin wrote: ↑Wed Dec 04, 2019 7:45 pm In our time, an average CPU runs Stockfish at 5 Mn/s.
So what's the point to stop Stockfish after 1000 nodes ?

The fact that stockfish needs 5 million NPS to make a move is disastrous enough,

Hi Peter,

kasinp wrote: ↑Wed Dec 04, 2019 8:21 pm I will take this opportunity to thank you for the tremendous job in creating (and maintaining) the invaluable CB-Emu collection.
It has helped me rekindle my love for computer chess and it is sure to be a source of joy for years to come. A true gem of a site!

thanks for your kind words, it's nice to see that people enjoy the emulations of these oldies.

(any chance we could see one of the Conchess models being emulated one day?)

Well, if we'll ever get the ROMs of these models, then yes of course.

Regards,
Franz

F.Huber wrote: ↑Wed Dec 04, 2019 1:33 pm

JohnW wrote: ↑Wed Dec 04, 2019 1:18 pm Just an fyi, you can download a lot of the Mephisto engines as uci if you wanted to run them on the same hardware to compare.

http://rebel13.nl/index.html

Or you could also download my CB-Emu program, which emulates more than 200 chess computers (not only Mephisto, but also Fidelity,
Novag, Saitek etc.), and it also includes MessChess for using all these devices as WB or UCI engines (e.g. in Winboard or Arena):
https://fhub.jimdo.com/

What an amazing thing!

To play a move or a game, a chess engine will get some hardware time.

The strength and "smartness" of an engine follow from the quality of the moves played. Whether to select these move it uses 1 nodes, 1K nodes, 1M nodes, 1B nodes... is irrelevant, the important part is how good of a result it can get with the available resources.

Putting a hardware time limit that is extremely low doesn't make sense either to assess an engine's strength, as anybody ever using it will be able to give the engine much more hardware time than the artificially low limit. How the engine's strength scale with more hardware time is much more relevant.

mclane's outlandish claims on this topic are notoriously wrong or irrelevant.

mclane wrote: ↑Wed Dec 04, 2019 9:06 pm

Vinvin wrote: ↑Wed Dec 04, 2019 7:45 pm In our time, an average CPU runs Stockfish at 5 Mn/s.
So what's the point to stop Stockfish after 1000 nodes ?

The fact that stockfish needs 5 million NPS to make a move is disastrous enough,

Just to add to your interesting observations - not all nps between the engines are the same. Back in the day, a great deal of focus in chess programming , was on the programming of chess smarts , the code usually found in evaluate.c. But it turned out, on much faster software, the largest Elo gains are not from evaluate.c , but are from search.c . It is the shape and depth of the search tree. I recently issued Weakfish, Weakfish has all of the evaluate code retained from stockfish and almost none of the search smarts. It is over 1000+ Elo weaker than Stockfish. The search smarts really kick butt somewhere between 5000/10000 nps to about 200,000 nps. At 200,000 nps, Stockfish is just about unbeatable by any human, only draws are given. From 200,000 nps to 5,000,000 nps, that is maybe a 500 Elo gain at best (from ~about 3000 Elo to about ~3500 Elo). If you take an engine like Crystal ,and play it against Stockfish at 1000 nps, it will be the stronger engine. If you play against it Stockfish at 200,000 nps, Stockfish comes up on top. So SF is not really a beancounter , but clearly it is not a "chess smart" engine either, it is however, really, really good at shaping the search tree. Not sure what one would call that, but that is where its strength comes from.

The search tree is wrong if you need so many positions to find the right moves.
So many of the evaluated positions are redundant.

Alayan wrote: ↑Wed Dec 04, 2019 10:13 pm To play a move or a game, a chess engine will get some hardware time.

The strength and "smartness" of an engine follow from the quality of the moves played. Whether to select these move it uses 1 nodes, 1K nodes, 1M nodes, 1B nodes... is irrelevant, the important part is how good of a result it can get with the available resources.

Putting a hardware time limit that is extremely low doesn't make sense either to assess an engine's strength, as anybody ever using it will be able to give the engine much more hardware time than the artificially low limit. How the engine's strength scale with more hardware time is much more relevant.

mclane's outlandish claims on this topic are notoriously wrong or irrelevant.

It’s Maybe irrelevant if you want to create a stupid chess engine, but if you want to create an intelligent chess engine it is not irrelevant.

Likewise, if you want to create an intelligent car, you are forbidden to use wheels. Legs are the way to go!

mclane wrote: ↑Thu Dec 05, 2019 6:17 am

Alayan wrote: ↑Wed Dec 04, 2019 10:13 pm To play a move or a game, a chess engine will get some hardware time.

The strength and "smartness" of an engine follow from the quality of the moves played. Whether to select these move it uses 1 nodes, 1K nodes, 1M nodes, 1B nodes... is irrelevant, the important part is how good of a result it can get with the available resources.

Putting a hardware time limit that is extremely low doesn't make sense either to assess an engine's strength, as anybody ever using it will be able to give the engine much more hardware time than the artificially low limit. How the engine's strength scale with more hardware time is much more relevant.

mclane's outlandish claims on this topic are notoriously wrong or irrelevant.

It’s Maybe irrelevant if you want to create a stupid chess engine, but if you want to create an intelligent chess engine it is not irrelevant.

Optimizing search (for hand written engines) has long proven to be the more intelligent (and efficient) choice to improve chess software.

The neural networks are a different beast, but they aren't hand-written either.

mclane wrote: ↑Thu Dec 05, 2019 6:15 am The search tree is wrong if you need so many positions to find the right moves.

But Mephisto NEVER finds the right moves! At least not in any reasonable time.

Like, how much time does Mephisto need to equal Stockfish at one second per move? 10 years per move? So where's its chess understanding? Until someone posts a position where Mephisto finds the right move while Stockfish *in one second* doesn't, all these claims of Mephisto having the "chess essence" seem empty.

Otherwise, you can easily modify Stockfish's source to make it count a node only every 10000 nodes, then Stockfish at 1000 nodes is going to be amazing and much faster than Mephisto.

The reason Stockfish is smart and Mephisto is dumb is that Mephisto needed stuff programmed into its evaluation that Stockfish can figure out on the fly, by searching more positions, because it has the time.