Too Much Deepness, Too Much Errors

[fern]

Looks like almost every engine today can goe very far in the tree search, but that is double edge sword. If evaluation is defective, each succesive ply will increase the mistakes. I have seen that in several engines. I wonder if this trend to go deeper and deeper has not gone too deep... A more balanced approach could be, I think, exception made you already have a very very good evaluation function, to come back a little to the old proven brute force approach, this time blessed with lot faster processors and techniques. Better to see 10 moves deep without leaving outside, due to defective eval, some important lines, than seeing badly 24 half moves forward.

Fern

[cdani]

As you sure know, is statistically better to see deeper than wider. I think is due to chess being more tactical than positional, at least in the matches between computers.

I'm sure than in fact chess is more positional than tactical, but this will be visible only when egines reach someting like 4000 elo, and will play very well positionally and with long term view, and at the same time don't make tactical mistakes.

[fern]

I am convinced that precisely because of the tactical nature of most chess, a wider look -but enough deep in any case- gives less room to misses.
Of course an old engine looking no more than 5 ply in middle game cannot compensate that shallowness with a wide search, but on the contrary if you look 20 moves and do not know well how to prune in order to avoid tactical mishaps, you will have them....
But of course you know lot more than me, I am just a player.

Fern
Just 10 ply deep regards

[cdani]

In some years the power of the computers will allow to test deeper. So for example one will be able to tune with 1 minute games what today requires 10 minute games.

So the values of the evaluation parameters will be different, and this will imply better long term moves.

If someone has 200 cpu at his disposal, he can advance a little this future for his engine with longer time test games.

[Engin]

its more common to search deeper with some kind of risky like human grandmasters do then to look only with brute force search every nonsense move thats waste of time. of course engine are missing something good moves here and there, but this algorithm is more human-like

[Vinvin]

Looks like almost every engine today can goe very far in the tree search, but that is double edge sword. If evaluation is defective, each succesive ply will increase the mistakes. I have seen that in several engines. I wonder if this trend to go deeper and deeper has not gone too deep... A more balanced approach could be, I think, exception made you already have a very very good evaluation function, to come back a little to the old proven brute force approach, this time blessed with lot faster processors and techniques. Better to see 10 moves deep without leaving outside, due to defective eval, some important lines, than seeing badly 24 half moves forward.

Fern

The problem is not the depth but the selectivity.
Stockfish and Komodo tends to prune the sacrificing lines that cause some surprises

[fern]

They are correlated in high degree as much you ned to select in order to prune. By definition, if you have NOt a perfect evaluation, there is room to leave aside what is important. On the contrary, if you have absolute wide serach yopu ate least avoid that problem...