Yet another time allocation heuristic

[sje]

Yet another time allocation heuristic:

This is one of those ideas I get which I suppose someone else may already using but if so, then I haven't read about it.

Here's an idea for modifying the basic search time allocation:

1) Have the program retain the predicted variations for all prior searches, both regular searches and ponder searches.

2) If the actual over-the-board played moves closely or exactly match the results of the prior N searches, then DECREASE the normal search allocation time.

3) If the actual over-the-board played moves FAIL to closely or exactly match the results of the prior N searches, then INCREASE the normal search allocation time.

This idea needs a nifty name for an inclusion in the wiki, but I can't think of one.

[hgm]

Did you test this in any way?

It sounds a bit like my method for saving fuel when driving at night: when the road had no curves for the past two miles, dim your head lights. For every 2 additional miles without curves, dim them further. After encountering a sharp turn, switch them back to full brightness.

(Needless to say, I don't have a drivers license! )

[sje]

Testing? Well, not much. I'd rather not take the fun away from someone else.

But it makes sense, though. If recent PVs strongly match over-the-board play, then the program's search + evaluation + trickery understands the position and so less time is needed for selection. And for the opposite, the opposite.

[sje]

In honor of HGM's post, I hereby name the above PV/play correlation trick:

The Headlight Heuristic

[hgm]