Couple more ideas

[Lyudmil Tsvetkov]

Yes, your points are quite valid. And still I wonder if, for example, the "hotspot bonus" idea might prove positive were it possible to test at more standard time controls. I mean, if the engine doesn't have time to find the "point" of the position you've told it to choose, of course that choice will fare poorly in your testing. But I guess I'm just repeating myself...

I'm running a local test of the latest variant at regular fishtest STC conditions right now. Again the VSTC looked good. So far at STC it is zero elo, but the error bars are still large.

If a patch depends on 'finding the point' you would expect it to scale up with longer TC, no?

It is possible that it scales with 60 sec, but does not scale with 15 sec.

Why?

Because at 9 sec the engine simply uses the above eval indications to achieve something else.

Again, I do not say this is the case, but it is possible.

Do you know how the patch performs at 3 sec per game? (just a rhetoric question)

Another thing is that the framework book contains 80 000 unique positions with reversed colours, with different 10 000-long portions of the book possibly displaying different or quite different structures/behaviour. SO, a 10 000 games test, no matter what the TC is, would never provide sufficiently accurate indication of what the trend is.

[Lyudmil Tsvetkov]

Interesting how this will perform on the framework.

A rule with almost no exceptions failing on the frameowrk - I would like to see that.

Well, maybe in the next 2 months or so.

Actually, this is the simplest closed positions rule, you can not think of a more simple rule - very simple and very efficient.

[vincenegri]

. I thought the hotspot idea failed at VSTC in your local testing. In any case, if my concern is a valid one, the idea should prove better as TC increases. If not, then maybe my concerns about fishtest are unfounded.

I can see why you might have got the wrong end of the stick. Hotspot passed at VSTC - in the sense that after 5,000 games the hotspot patch had a 95% LOS over master. But on looking in depth at the actual games, the extra wins were not coming about as a result of king-side attacks based on the e5/g5 bind.

Of course it's possible that the eval term was having some other beneficial effect. For example, Lyudmil's en passant bonus was successful at STC and LTC in the framework, but another patch also went through at the same time that overlapped and 'soaked up' whatever it was the en passant bonus was measuring.

Anyway, as it stands right now hotspot is behind in the test. Still within error bars, but it doesn't look like it will perform as well as it did at VSTC. And it wouldn't be the first patch to act like this (do better at low depths) if that's how it turned out.

Coming back to fianchettos, I think the same considerations apply here.
joergoster's patch (http://tests.stockfishchess.org/tests/v ... 02160ebec5) simply gave a bonus to possessing a k-side fianchetto. But forcing the engine "with a hammer" like this is unlikely to work because if it can't handle the bishop properly, the fianchetto will just be a detrimental thing. And that's what happened; the patch didn't just 'not pass' but lost Elo. You can argue about opening books when a patch fails to show gain, but not when it is losing Elo.

In the current way an engine looks at the position, there are a lot of reasons for it to superficially dislike a king-side fianchetto. Here's a pretty canonical KID arrangement:

[d]5rk1/5pbp/3p1np1/3Pp3/4P3/8/8/8 w - - 0 1

So it has moved a pawn in front of the king ("bad"), and its bishop has mobility of 2. Even if the knight moves, it is impeded by its own e-pawn, which is blocked.

Now practical experience tells us that these are not real problems. But the engine does not learn by experience (at least, learning is always turned off in tournaments and tests) so it only has what we have given it to go on. And until it knows how to use a fianchettoed K-bishop, it will see it as a liability.

btw, you'll notice that I wrote 'practical experience tells us' above. That's because I reserve judgement about the ultimate truth of the question. Computers have upended some holy cows in chess already, and it is conceivable that in years to come we will find that certain structures which are excellent in practical play between humans are not all we cracked them up to be. Certainly we should avoid taking things for granted, because engines have accelerated the process that began in the 40s and was brilliantly described by Watson wherein chess has become ever more concrete and "rule-independent".

[Lyudmil Tsvetkov]

. I thought the hotspot idea failed at VSTC in your local testing. In any case, if my concern is a valid one, the idea should prove better as TC increases. If not, then maybe my concerns about fishtest are unfounded.

I can see why you might have got the wrong end of the stick. Hotspot passed at VSTC - in the sense that after 5,000 games the hotspot patch had a 95% LOS over master. But on looking in depth at the actual games, the extra wins were not coming about as a result of king-side attacks based on the e5/g5 bind.

Of course it's possible that the eval term was having some other beneficial effect. For example, Lyudmil's en passant bonus was successful at STC and LTC in the framework, but another patch also went through at the same time that overlapped and 'soaked up' whatever it was the en passant bonus was measuring.

Anyway, as it stands right now hotspot is behind in the test. Still within error bars, but it doesn't look like it will perform as well as it did at VSTC. And it wouldn't be the first patch to act like this (do better at low depths) if that's how it turned out.

Coming back to fianchettos, I think the same considerations apply here.
joergoster's patch (http://tests.stockfishchess.org/tests/v ... 02160ebec5) simply gave a bonus to possessing a k-side fianchetto. But forcing the engine "with a hammer" like this is unlikely to work because if it can't handle the bishop properly, the fianchetto will just be a detrimental thing. And that's what happened; the patch didn't just 'not pass' but lost Elo. You can argue about opening books when a patch fails to show gain, but not when it is losing Elo.

In the current way an engine looks at the position, there are a lot of reasons for it to superficially dislike a king-side fianchetto. Here's a pretty canonical KID arrangement:

[d]5rk1/5pbp/3p1np1/3Pp3/4P3/8/8/8 w - - 0 1

So it has moved a pawn in front of the king ("bad"), and its bishop has mobility of 2. Even if the knight moves, it is impeded by its own e-pawn, which is blocked.

Now practical experience tells us that these are not real problems. But the engine does not learn by experience (at least, learning is always turned off in tournaments and tests) so it only has what we have given it to go on. And until it knows how to use a fianchettoed K-bishop, it will see it as a liability.

btw, you'll notice that I wrote 'practical experience tells us' above. That's because I reserve judgement about the ultimate truth of the question. Computers have upended some holy cows in chess already, and it is conceivable that in years to come we will find that certain structures which are excellent in practical play between humans are not all we cracked them up to be. Certainly we should avoid taking things for granted, because engines have accelerated the process that began in the 40s and was brilliantly described by Watson wherein chess has become ever more concrete and "rule-independent".

Not true, Vince.

Computers have tactical, short-term power, they lack strategic, long-term power.
Tactics make them strong, much stronger than humans, but they are way behind in strategic handling of the game.

Tactics means good search, strategy means good eval.
Eval is knowledge, strategy is eval.
Knowledge is well-tuned eval and more available reasonable eval terms. The better tuned your eval terms are, the more knowledge you possess; the more reasonable eval terms you have, the more knowledge you have.

I bet that if SF knew about kingside fianchettoe, and granted it good search, it would fianchettoe its bishop on the king side every second game; actually, I think, even more frequent than that.

The problem with Joerg's patch is not that SF does not recognise the importance of kingside fianchettoes, when they are available; I tried this many times, and, whenever I fianchettoe the king bishop instead of SF, it very soon gives that line as better than other lines it likes to choose. That repeats time and again, it sees kingside fianchettoes are good when they are available in most cases after a search, but the problem is how to make the engine do this.

As I pointed out in another thread, the problem with king fianchettoe in SF is interference with king safety terms, quite probably the maximum king safety bonus; 3 pawns in front of the king give this bonus, but a h2.g3.f2 pawns with Bg2 do not. So, even if bonus is given, the engine will tend to avoid playing in this way and the patch will fail simply because they are not sufficent number of fianchettoe games.

The KID is a different story, besides the fianchettoed bishop you msut also see a couple of other important terms, like pointed chain, but a kingside fianchettoe would be valid and the optimal solution for a very wide range of openings.

So again, you should not mix 2 things: engines are tactically/search-based strong, but positionally/strategically/eval-based weak. Even the top ones.

[Lyudmil Tsvetkov]

Btw, Vince, just to ask you: did you try implementing the hotspots rule as a single available pawn structure with closed center: d4,e5,g5 white pawns, d5,e6,f7,g6 black pawns?

I think this might be essential for the success of the patch.

An open center or lack of a backward pawn on f7 might lead you nowhere.

[vincenegri]

Btw, Vince, just to ask you: did you try implementing the hotspots rule as a single available pawn structure with closed center: d4,e5,g5 white pawns, d5,e6,f7,g6 black pawns?

I think this might be essential for the success of the patch.

An open center or lack of a backward pawn on f7 might lead you nowhere.

Not yet - largely because the tighter you make the definition, the less frequently the position arises in search and the longer the test becomes.

It is next in the queue.

[Lyudmil Tsvetkov]

Btw, Vince, just to ask you: did you try implementing the hotspots rule as a single available pawn structure with closed center: d4,e5,g5 white pawns, d5,e6,f7,g6 black pawns?

I think this might be essential for the success of the patch.

An open center or lack of a backward pawn on f7 might lead you nowhere.

Not yet - largely because the tighter you make the definition, the less frequently the position arises in search and the longer the test becomes.

It is next in the queue.

But this one will suceeed.

[vincenegri]

The problem with Joerg's patch is not that SF does not recognise the importance of kingside fianchettoes, when they are available; I tried this many times, and, whenever I fianchettoe the king bishop instead of SF, it very soon gives that line as better than other lines it likes to choose.

But you will have 'forced' SF to fianchetto in this way when you realise that it is appropriate to do so, picking a good time and the right board situation. That is evidenced by the line subsequently having a good eval.

The king-side fianchetto is not always the right move - if it were, all GMs would open with 1.g3 g6 2. Bg2 Bg7

So when forcing SF, you act as a very sophisticated additional eval term.

Putting in a brute force 'fianchettos are good' rule doesn't work because it will end up fianchettoing too early by rote. For example, you don't want to end up playing e4 e5 Nf3 Nc6 Bb5 g6 (smyslov) as this line is under a theoretical cloud.

[Lyudmil Tsvetkov]

The problem with Joerg's patch is not that SF does not recognise the importance of kingside fianchettoes, when they are available; I tried this many times, and, whenever I fianchettoe the king bishop instead of SF, it very soon gives that line as better than other lines it likes to choose.

But you will have 'forced' SF to fianchetto in this way when you realise that it is appropriate to do so, picking a good time and the right board situation. That is evidenced by the line subsequently having a good eval.

The king-side fianchetto is not always the right move - if it were, all GMs would open with 1.g3 g6 2. Bg2 Bg7

So when forcing SF, you act as a very sophisticated additional eval term.

Putting in a brute force 'fianchettos are good' rule doesn't work because it will end up fianchettoing too early by rote. For example, you don't want to end up playing e4 e5 Nf3 Nc6 Bb5 g6 (smyslov) as this line is under a theoretical cloud.

The fact that SF picks up another move, but when fianchettoe is played, it decides in a couple of moves this is better, tells volumes about kingside fianchettoe.

There is almost no exception to the fianchettoe rule - Fischer used to fianchettoe kingside in more than half of his games. You can do so in practically every opening. The Smyslov line in the Ruy Lopez is seen as bad by engines even after a short search, things like that happen, but there are many other openings where king fianchettoes are not in vogue, but where they could easily be introduced with benefit.

Not all GMs are equally strong, look at the games of the top of the top.

With SF, the problem is this: it gives max king safety bonus for h2,g2,f2, when g3 advances one square, maybe some 20cps penalty is given. Could you calculate what happens when you give for king fianchettoe lower than 20cps bonus?
So I think the solution is to consider max king safety bonus also in the case of h2,g3,f2 pawns, when giving the bonus.

Only problem if SF does not apply this is it loses 35 elo or so.

It seems only Daniel understands me on this forum...

[vincenegri]

Fischer used to fianchettoe kingside in more than half of his games.

Only when black. Pulling up Fischer's games in a database:

As white: 558 games, of which 66 feature a king-side fianchetto (defined as a bishop on g2 and a pawn on g3 on or before move 15)

As black: 396 games, of which 236 feature a king-side fianchetto.

Bobby sure loved his KID

Of course this is due to Bobby being a 1.e4 guy… so perhaps the issue is that one should only fianchetto in closed/semi-closed openings.