What separates the top engines from the rest?

[fern]

".....it just might see or might not see the consequences of such a move. If the evaluation function gives a higher score at the end of the given line than it gave on any other lines it will play the move..."

But, to that, to "evaluate" as good or bad, to evaluate consequences, something else than just material score must be present, isnt so? Or consequences at last are always material or a mate?

I am sure that the engine does not know what he is doing, it is not a sentient entity, but maybe it is in fact doing some positional evaluation.
Please, clear this to me.

fern

[voyagerOne]

Simple.

Mastering the art of pruning to look at sharp tactical lines.

[fern]

Your response only re word the issue.
I still can ask how that sharp lines are discovered IF there is not clear material advantage. Or how you dismiss material advantage that takes you to doom?
Somewhere some other principles must be operating.

fern

[Don]

".....it just might see or might not see the consequences of such a move. If the evaluation function gives a higher score at the end of the given line than it gave on any other lines it will play the move..."

But, to that, to "evaluate" as good or bad, to evaluate consequences, something else than just material score must be present, isnt so? Or consequences at last are always material or a mate?

I am sure that the engine does not know what he is doing, it is not a sentient entity, but maybe it is in fact doing some positional evaluation.
Please, clear this to me.

fern

Hi Fern,

The good programs finally can evaluate compensation for a pawn, compensation for the exchange and even more. And they are often not too far off the mark. They are not perfect, but they are quite good in that respect compared to a few years ago. If you put a gambit in Komodo it will believe the position is about equal, in marked contrast to most strong programs of 30 years ago. Also, any program with good king safety can recognize several pawns worth of compensation for king safety.

Whether human or computer the "skill" is to appreciate the strengths or weakness of a position without actually seeing "concrete compensation." For example giving up a couple of pawns for the attack even though you cannot see the checkmate or even the direct recovery of material. Programs are starting to be able to do that, humans have always been pretty good at it.

Computers have closed some of the gap for 2 reasons. One reason is that they search so deeply that they will sometimes see outright material compensation for what we would call a "positional" sacrifice. The other reason is because the evaluation has improved to the point that if a sacrifice is called for, computers are usually "sympathetic" to the move. In the old days such moves might be near the end of the move list so to speak. I know that with Komodo (as well as other strong programs) there is often not much difference in score between the sacrifice move and some other, even at very shallow depths. If a deeper search makes it more obvious they will often play the needed move.

There is still much work to be done, but we have seen a lot of progress. I think a lot of the younger programmers here take "good" evaluation for granted - they lack historical perspective. It's no big deal to them that programs are doing these things now and they don't see how far we have come. This tends to make them underestimate the role of the evaluation function. It's like being born wealthy and not fully appreciating where the money came from.

[fern]

Hi Don:
In fact, although I posted as making a question, I was pretty sure things were as you described them The sheer use of the word "evaluation show that. Even old dedicated units as Superconny knew how to sacrifice a bishop in h7. Maybe in this case was a very scripted sequence, but no matter what, we can call it positional maneuver. Also we, human, make many moves on the ground of a pre learned protocol, almost automatically, but that does not prevent we can talk of positional acumen, etc.
The subjective content of the process of choosing a move does not matter in human, even less in comps. So if a code can discriminate nuances about positions, surely he is positionally playing. I can see that clearly in Komodo for certain.

still waiting your great GUI revolutions regards
Fern