tuning for the uninformed

[Robert Pope]

Hi Folkert,

The sigmoid function is used to map the score of the evaluation (which is f.i. in the range of -100 to +100) to a value in the range of 0 to 1.
A score of -infinity results in a value of 0, a score of +infinity in a value of 1. A score of 0 results in a value of the sigmoid of 0.5.

So instead I could also do (score + 10000) / 20000 ? (assuming eval score range is -10k ... 10k)

Sigmoid is much better because the shape recognizes that it is the scores near 0 that are most important to differentiate. Going from +1 to +4 is much more significant than going from +10 to +14.

[hgm]

But not if you actually want to be able to win in a +10 position...

[Joost Buijs]

That is why we need search, it is very likely that with an advantage of +10 (I assume 10 pawns) the check-mate will be in the horizon of a decent search. Nowadays most engines can find a check-mate in (lets say 20 moves) in just a few seconds.

[jdart]

Some King attacks are still not trivial for programs to execute, especially when there are quiet moves (zwischenzuegen) in the PV, or one or more sacrifices are involved.

--Jon

[Joost Buijs]

Some King attacks are still not trivial for programs to execute, especially when there are quiet moves (zwischenzuegen) in the PV, or one or more sacrifices are involved.

--Jon

Programs that prune less are usually better at these type of positions. There clearly is a difference between finding the check-mate in a position with an already +10 material advantage or finding a combination that leads to mate in a more or less equal position.

Having a very good evaluation of king-safety really helps, also with pruning decisions, this is an area where NN type engines probably shine. Unfortunately this is where my engine is very bad at.