Instead of

3 + depth / 4

We now have:

3.2 + depth / 3.8

The integer defaults are almost the same and of course easier, but test passed....

@zamar @glinscott: can you have a look ? I would like to test the 3 parameter tuned model against the new master: http://tests.stockfishchess.org/tests/view/54a895570ebc597e33447ece

@mcostalba: yes, it's called a tuning patch. you can't expect the tuned values to be very far from the initial ones, because the initial ones were already good. Note that I used SPRT(-0.5,4.5)+SPRT(0,5) to reduce the risk of regression (type I error), while leaving a sufficiently small resolution (5 bayeselo) to detect a small elo gain (type II error).

It is a tuning patch and passed [0, 5], so

Approved.

@lucasart: You can test the 3 parameters version against 2 parameters even before 2 param version is committed.

However I see one version where you hard coded max reductions to 10 plies. This is risky and I wouldn't recommed you doing that... The rule will kick in somewhere around depth 30, so we have no way of measuring the impact and it'll only affect very long time controls.

Looks good!