Computer Chess Club Archives




Subject: Re: What to do with Horizon effect?

Author: Amir Ban

Date: 15:03:21 11/01/98

Go up one level in this thread

On November 01, 1998 at 13:41:20, Roberto Waldteufel wrote:

>On November 01, 1998 at 10:20:02, Amir Ban wrote:
>>On October 31, 1998 at 17:50:48, Robert Hyatt wrote:
>>>not much you can do...  horizon effect happens anytime you stop the search
>>>in a non-quiet position...  everybody sees it...  just get faster.. :)
>>I don't agree with this statement. The exposure to horizon effects is what
>>usually limits a program's strength, more than the depth it can reach. The fact
>>that it cannot be totally overcome is irrelevant to the fact that limiting its
>>damage is a first priority. Just getting generally deeper is a poor way to do
>>that, and not very effective either.
>So what do you do to "limit the damage" of the horizon effect. Do you perform
>any special quiescence tests other than following captures/pawn promotions and
>perhaps a few checks beyond the main search? Whereas it may be a crude solution
>to try to "outrun" the horizon with fast searching, it is the only practical
>solution of which I am aware. Greater depth makes it less likely, but never
>impossible, for a horizon misevaluation to propogate back to the root.

It's not possible to get rid of the horizon effect completely, but that's not
the target. The target is to lower the effect, and by that to make the program
stronger. If it's free, the more the better. Practically, there's a tradeoff
involved, and at some point you're better off accepting some horizon effect
rather than paying the price to limit it further. Most of us are probably still
far from this break-even point so we should keep trying.

The horizon effect should be handled through your extension and quiescence
methods. If you know how to search deeper, by all means do it, but that's a poor
remedy for the horizon effect. In fact, a program that handles the horizon
effect well will appear to outsearch a program that doesn't, other things being
equal, which means that it makes sense to sacrifice depth to handle the effect

Regarding your question about quiescence the answer is no. I don't think the
right way is to add more stuff to the quiescence search, and I suspect that many
programs try too hard in quiescence. This can be not only expensive but
counterproductive too, since it is quite possible for a program to become
stronger by seeing less, a consequence of the horizon effect.


This page took 0 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.