Coming soon - SF-McBrain 9 ...

[carldaman]

+1 It is a vastly superior method to the hash dump everyone else is using these days.

Which is why I'm surprised to see that fork under-maintained.

There appears to be enough interest, certainly from the more savvy users. I'm very surprised the top commercial engines have not implemented something similar. Komodo's authors have always said they'd do it if there is "sufficient interest", but how do you define that? They can just say there's isn't that much interest, in their view.

A very strange situation has arisen, where a most useful feature is totally ignored or sidestepped by the programmers. One can hope a new SF fork could be developed that includes learning, but we can't really expect people to work for free. That's why you'd think the commercial authors would jump at this, but sadly it appears almost hopeless.

If I was a commercial author trailing SF by a serious margin, I couldn't wait to add features that others do not have, and market the product around that.
No one thinks like that any longer. Chasing Elo now trumps even common sense.

[Ovyron]

No one thinks like that any longer. Chasing Elo now trumps even common sense.

But not even open source engines like SF are having those features!

What's the point of having Open Source with features people want if nobody is implementing it?

So here we have 3 different Learning methods fully implemented by people:

Result learning by Michael Sherwin.
Self-learning by Kelly Kinyama.
Persistent Hash by Jeremy Bernstein.

But nobody capable of implementing any of those into Stockfish seems interested...

[MikeB]

The release version of McBrain v9 , based on SF-9 is final. While I wait for some binaries to be compiled, you can follow along on this 2000 game match SF-9 vs McBrain v9.

Tc 60 Seconds plus 2 second increment, using Kai's 3 move openings , starting at position #24 , hence first round in the pgn file will be labeled 24 and not 1.
Pondering is off.
12 concurrent games on a 2010 12 Core Mac Pro

Result file will be updated every 6 games or roughly every 4 to 6 minutes as will the pgn file. This match will run about 20 hours.

Results: https://www.dropbox.com/s/qtrti9qx494yf ... a.txt?dl=1


Pgn File: https://www.dropbox.com/s/kbeqd2o608p22y4/all.pgn?dl=1

[carldaman]

No one thinks like that any longer. Chasing Elo now trumps even common sense.

But not even open source engines like SF are having those features!

What's the point of having Open Source with features people want if nobody is implementing it?

So here we have 3 different Learning methods fully implemented by people:

Result learning by Michael Sherwin.
Self-learning by Kelly Kinyama.
Persistent Hash by Jeremy Bernstein.

But nobody capable of implementing any of those into Stockfish seems interested...

It would take a lot of work to maintain such a SF fork and most people don't want to work for free. That's why it would make more sense for a strong commercial engine to incorporate such a feature. I'd pay extra for such a product, Elo gains or not, wouldn't you?

[Ozymandias]

No one thinks like that any longer. Chasing Elo now trumps even common sense.

But not even open source engines like SF are having those features!

What's the point of having Open Source with features people want if nobody is implementing it?

So here we have 3 different Learning methods fully implemented by people:

Result learning by Michael Sherwin.
Self-learning by Kelly Kinyama.
Persistent Hash by Jeremy Bernstein.

But nobody capable of implementing any of those into Stockfish seems interested...

It would take a lot of work to maintain such a SF fork and most people don't want to work for free. That's why it would make more sense for a strong commercial engine to incorporate such a feature. I'd pay extra for such a product, Elo gains or not, wouldn't you?

True, you can't expect people to work for free, but the reality if it... is that they do. Hence my surprise about the absence of a properly maintained fork, among the myriad of SF forks, highlighting such a feature. Why having a SF fork, if it's going to be lost in the sea (pun intended)? Why not go with one feature that will instantly make it shine?

As for commercial developers, I expect to see them turn to features, as soon as Elo chasing starts giving decreasing benefits.

[Ovyron]

I'd pay extra for such a product, Elo gains or not, wouldn't you?

Who knows, I didn't get Komodo 11, but I don't know if Learning would have changed my mind, definitively doesn't seem worth the price tag if it's basically K10 with learning.

The sad truth is, that Vasik Rajlich tried to add learning (Persistent Hash) to Rybka, and his sales didn't increase, so it was removed in a later version. Robert Houdart tried to add learning to Houdini, and his sales didn't increase, so it was removed in a later version. Shredder 10 had it (it actually had better implementations than the ones discussed on this thread), but its sales didn't increase, so it was removed in a later version.

Learning just has the poorest stories when it comes to commercial software. I even helped Ubaldo Andrea Farina to implement Learning in Chiron, which awarded me lifetime free updates, but apparently he saved his time implementing it and removing it from a later version, by not putting it there at all in the first place! XD

So people's dreams are related to strong engines like Stockfish getting some kind of learning, but when it comes to the top commercial engines, they will sell basically the same if they implement Learning, or just improve 20 elo, and so, they spend their time in the chase, knowing the audience they already have will buy it after that effort.

[Ozymandias]

I don't know if Learning would have changed my mind, definitively doesn't seem worth the price tag if it's basically K10 with learning.

The sad truth is, that Vasik Rajlich tried to add learning (Persistent Hash) to Rybka, and his sales didn't increase, so it was removed in a later version. Robert Houdart tried to add learning to Houdini, and his sales didn't increase, so it was removed in a later version. Shredder 10 had it (it actually had better implementations than the ones discussed on this thread), but its sales didn't increase, so it was removed in a later version.

[...]

when it comes to the top commercial engines, they will sell basically the same if they implement Learning, or just improve 20 elo, and so, they spend their time in the chase, knowing the audience they already have will buy it after that effort.

I, on the other hand, bought H4 based on the new syzygy support, and not because of the meagre Elo increase over H3. It will depend on a personal basis.

Rybka's Persistent Hash usefulness was completely tied to the Rybka version that generated it. Once a sufficiently improved version comes along, such work becomes worthless. Other implementations lacked timing and/or a good marketing strategy. In the wake of the AlphaZero match, any good implementation (the first two you mentioned earlier, for example) with appropriate marketing, should get a desirable "development time/increased income" ratio.

Your last point has been true up to now, but given the recent event I just mentioned and the difficulty programmers might encounter, when it comes to obtaining significant Elo increases, it could stop being true at any moment.

[Ovyron]

Once a sufficiently improved version comes along, such work becomes worthless.

But this is true for ALL engines. The Stockfish version Pre-Contempt gives completely different eval than the Post-Contempt version. Imagine that someone implemented Position Learning into the PreContempt version and then ALL analysis would have been tied for it and would have become worthless after the introduction of Contempt because the engine would have preferred radically different variations.

And this time old versions are obsolete, back in the time of Stockfish 2 was introduced, I could use different versions of Stockfish when this happened, and enjoy low redundancy as if they were completely different engines.

Nowadays my analysis from November 2017 is obsolete, and with it all the analysis of the past 10 years, but no amount of learning or engine features would have saved it, so "all your chess work will become obsolete soon" is something people have to learn to live with, and it's something impossible to stop in any learning implemented.

Your last point has been true up to now, but given the recent event I just mentioned and the difficulty programmers might encounter, when it comes to obtaining significant Elo increases, it could stop being true at any moment.

If you're the same Ozymandias I know, I don't need to tell you this, but you'll remember this has been dreaded for the last 13 years, and it never happens. Rybka 1.0 happened and then Rybka 3 happened, and then the whole Clone Wars happened and Houdini became the top clone, and Doch happened and after Don Dailey died Komodo became top engine for a bit, and then Stockfish gained 400 ELO and my neighbor and her dog each got their own Stockfish Branch that treats Contempt differently than dev by default...

Alpha Zero may have let us peek into the future of how things will be done, where programmers don't bother teaching their engines material values or check extensions at all, or it may be the next Deep Blue only remembered by some feat in a time when Crippled Stockfish 8 gets defeated by 72-0 by the top engine of the time, but the imminent scarcity of ideas to improve chess engines, where programmers find nothing else to do so they finally implement features that users want, may not be witnessed by my granddaughters...

[Ozymandias]

Once a sufficiently improved version comes along, such work becomes worthless.

But this is true for ALL engines. [...] no amount of learning or engine features would have saved it

Your last point has been true up to now, but given the recent event I just mentioned and the difficulty programmers might encounter, when it comes to obtaining significant Elo increases, it could stop being true at any moment.

the imminent scarcity of ideas to improve chess engines, where programmers find nothing else to do so they finally implement features that users want, may not be witnessed by my granddaughters...

If I understood Sherwin's implementation correctly, the basic evaluation is that of the engine currently analysing, past experience isn't vinculating unless it has become overwhelming.

Rybka 1 meant a 150 Elo jump, in one go, do you know how many years and different versions, among the top 3 engines, has taken to accumulate the latest 150 Elo points? About three years.

[carldaman]

No one thinks like that any longer. Chasing Elo now trumps even common sense.

But not even open source engines like SF are having those features!

What's the point of having Open Source with features people want if nobody is implementing it?

So here we have 3 different Learning methods fully implemented by people:

Result learning by Michael Sherwin.
Self-learning by Kelly Kinyama.
Persistent Hash by Jeremy Bernstein.

But nobody capable of implementing any of those into Stockfish seems interested...

It would take a lot of work to maintain such a SF fork and most people don't want to work for free. That's why it would make more sense for a strong commercial engine to incorporate such a feature. I'd pay extra for such a product, Elo gains or not, wouldn't you?

True, you can't expect people to work for free, but the reality if it... is that they do. Hence my surprise about the absence of a properly maintained fork, among the myriad of SF forks, highlighting such a feature. Why having a SF fork, if it's going to be lost in the sea (pun intended)? Why not go with one feature that will instantly make it shine?

As for commercial developers, I expect to see them turn to features, as soon as Elo chasing starts giving decreasing benefits.

Many people do work for free, of course, but they get to choose what to work on and set their own priorities... It's not easy to get them to do one's bidding that way.