Modern Computer Chess Theory: Ply vs Rating limits

[CRoberson]

In the past, several things were speculated:
o Search is everything....
o Research in eval is of lesser value
o 14 to 16 ply is all that is needed. After that, more is unnecessary.

Here is my question: which I believe could be researched to some degree:

Beyond what ply does increased search provide little improvement? Another way to ask it is: How much difference does 1 ply make at increasing depths?

For instance, do modern computers and programs using 24 hours per move produce nothing but draws? I am assuming both programs don't use EGBBs or EGTBs.

Also, does a two program rating gap of R from G/120 transfer to games at say 24 hours per move?

Using ICC or some server, several of us could test this.

[Daniel Shawul]

o Search is everything....

I think not. Strategy is still important but yes search is much
more important than eval

o Research in eval is of lesser value

True. Relatively that is.

o 14 to 16 ply is all that is needed. After that, more is unnecessary.

Disagree. We still need deep searches in pawn endgames, blocked positions or any other position where the long-rangers are gone. Maybe you should qualify the question 14-16 ply in the middle game which I am sure is what you had in mind.

I believe more plies more elo. There is definitely a flattening of the rate of return per ply but I think 8x8 board is too wide for just 14-16 _half_ plies. Chess on 3x3 has long tactics if I read the current threads correctly.

[Don]

In the past, several things were speculated:
o Search is everything....
o Research in eval is of lesser value
o 14 to 16 ply is all that is needed. After that, more is unnecessary.

It was said in the 70's that after 7 or 8 ply we would not expect much improvement from searching deeper. It sounds stupid now but some people believed it back then.

Here is my question: which I believe could be researched to some degree:

Beyond what ply does increased search provide little improvement? Another way to ask it is: How much difference does 1 ply make at increasing depths?

There is no magic number, there is a very gradual falling off and it's very smooth. If you were doing a 25 ply search a 26 ply search would beat it and it would not take very many games to prove this statistically.

If you were doing 40 ply searches you probably would have to play a LOT of games to prove 41 ply is better, but it would be better.

For instance, do modern computers and programs using 24 hours per move produce nothing but draws? I am assuming both programs don't use EGBBs or EGTBs.

No. Computers are several orders of magnitude away from playing good chess. I have speculated that they are well over 1000 ELO away from perfect chess - but of course this is only a wild guess.

Also, does a two program rating gap of R from G/120 transfer to games at say 24 hours per move?

The correlation is probably very high. You cannot test this unless you run at least several hundred games and that would only give you a rough estimate.

Using ICC or some server, several of us could test this.

[hgm]

I think it all depends on how clever the eval is. Positional gains in the long run translate themselves into material gain. So even without any positional knowledge in eval, searching very deep will bring the devastating consequences of positionally weak moves within the horizon, and therefore cause selection of positionally good moves. The further you can look ahead, the better your resolution gets on the positional scores.

OTOH, an engine with a very good positional eval will already have a good resolution of those scores at low depth, and will not benefit nearly as much from extra depth as one with poor positional knowledge. It ony gains the occasional correction of the exceptional cases where its general positional knowledge happens to fail, but in most cases would have no difficult toplay good moves at low depth.

[Uri Blass]

I think it all depends on how clever the eval is. Positional gains in the long run translate themselves into material gain. So even without any positional knowledge in eval, searching very deep will bring the devastating consequences of positionally weak moves within the horizon, and therefore cause selection of positionally good moves. The further you can look ahead, the better your resolution gets on the positional scores.

OTOH, an engine with a very good positional eval will already have a good resolution of those scores at low depth, and will not benefit nearly as much from extra depth as one with poor positional knowledge. It ony gains the occasional correction of the exceptional cases where its general positional knowledge happens to fail, but in most cases would have no difficult toplay good moves at low depth.

You may be right for very big depths but I believe that with the depths that engines get today it is not the case.

It is possible to test and find that only material evaluation earn less elo from additional ply relative to only piece square table evaluation in depths that you practically can test.

The same also happens when you compare a good evaluation with only piece square table evaluation and I stongly believe that the programs with a good evaluation earn more from increasing the depth in depths that we can achieve practically.

I can add that without positonal knowledge searching very deep is not going to help programs to choose the right drawing move.

Even with very big depth
programs may start with moves like 1.a4 because it also does not lose material even if you search into depth of 100 plies and playing good moves in bad positions may help them not to lose against weak engines but they will often make draws against them.

[Laskos]

In the past, several things were speculated:
o Search is everything....
o Research in eval is of lesser value
o 14 to 16 ply is all that is needed. After that, more is unnecessary.

It was said in the 70's that after 7 or 8 ply we would not expect much improvement from searching deeper. It sounds stupid now but some people believed it back then.

Here is my question: which I believe could be researched to some degree:

Beyond what ply does increased search provide little improvement? Another way to ask it is: How much difference does 1 ply make at increasing depths?

There is no magic number, there is a very gradual falling off and it's very smooth. If you were doing a 25 ply search a 26 ply search would beat it and it would not take very many games to prove this statistically.

If you were doing 40 ply searches you probably would have to play a LOT of games to prove 41 ply is better, but it would be better.

For instance, do modern computers and programs using 24 hours per move produce nothing but draws? I am assuming both programs don't use EGBBs or EGTBs.

No. Computers are several orders of magnitude away from playing good chess. I have speculated that they are well over 1000 ELO away from perfect chess - but of course this is only a wild guess.

Also, does a two program rating gap of R from G/120 transfer to games at say 24 hours per move?

The correlation is probably very high. You cannot test this unless you run at least several hundred games and that would only give you a rough estimate.

Using ICC or some server, several of us could test this.

The basic estimate is rather empirical and not so wild as one could imagine. In Marco's and Zach's _widening_ tree model (correct, rather than the wrong, Bob's purely exponential tree model), the search by itself to a certain ply would be a good indicator of strength for the future, with the present day eval, because for shallower depths on the tree, the tree anyway gets wider and wider with increasing _target_ depth, compensating for weak eval.

Therefore, empirically, let's assume that one ply at ply 10 gives 90 Elo points, 1 ply at ply 20 gives 70 points. These are verifiable numbers, close to reality for many engines.
Thus the geometric decrease in the Elo gain for 1 additional ply is given by (Log(90/70))/10 ~ 0.02513.
With this factor, assuming 20 ply depth of modern engines and 70 Elo points gain for additional ply, the total increase for "perfect" engine in the _search_ is 70/0.02513 ~ 2800 Elo points. This engine might still not be a "perfect" engine due to a weak eval, this is the bottom value of gain in strength for a "perfect" engine compared to today's say Stockfish (or a bottom of 6000 Elo points strength taking Stockfish as 3200).
As for 40 ply to 41 ply, the difference would not be extremely small, somewhere 70/(1.02513)^20 ~ 43 Elo points, measurable with 95% confidence in ~200 games, not a LOT as you said.
Also, the present day top engines would have 1 part per million to have a draw in a game with this almost "perfect" engine, even if the game of chess is really a draw. All the rest are won by this almost "perfect" engine.
If improvement in engines is ~60 Elo points per year, then there are at least 2800/60 ~ 47 years for progress in chess, and a game for programmers

Kai