Page 1 of 1

Methods to stably evaluate nodes?

Posted: Fri May 04, 2007 12:18 pm
by smrf
Especially at MP methods the node evaluations seems not to be determenisitic, because the result depends on the sequence of walking the tree.

Thus I believe that some programmers do not even store such values into the transposition table, because one could never rely exactly on those figures.

Are performant solutions known to get stable evaluations despite of any calculating sequence? I am just about to "invent" an alternative, but it could not be bad to learn about existing alternatives.

Regards, Reinhard.

Re: Methods to stably evaluate nodes?

Posted: Sat May 05, 2007 5:14 am
by smrf
Hmm, I understand, that there is no interest in discussing MP stable node evaluations. Thus I will shut up concerning my solution.

Re: Methods to stably evaluate nodes?

Posted: Sat May 05, 2007 5:33 am
by Uri Blass
smrf wrote:Especially at MP methods the node evaluations seems not to be determenisitic, because the result depends on the sequence of walking the tree.

Thus I believe that some programmers do not even store such values into the transposition table, because one could never rely exactly on those figures.

Are performant solutions known to get stable evaluations despite of any calculating sequence? I am just about to "invent" an alternative, but it could not be bad to learn about existing alternatives.

Regards, Reinhard.
I do not understand what is the meaning of MP methods.
Do you talk about multi-processor?

Note that the evaluation of movei is dependent on the path and not only on the leaf position but I do not support more than one processor and I do not plan to do it in the near future.

My evaluation has 2 parts(one static part that is dependent only on the position that I evaluate and one part that is dependent on the last game moves before the position)

I think that there are not many programs that have an evaluation function that is dependent on the path.

Uri

Re: Methods to stably evaluate nodes?

Posted: Sat May 05, 2007 5:43 am
by smrf
Hi Uri,
... I think that there are not many programs that have an evaluation function that is dependent on the path. ...
any method driven by the nature of previous moves is that unstable. And in MP approaches that is becoming obvious. Maybe in SP approaches it will rarely be noticed, but also there a node revisited by another possible previous node (which could lead to the same child node by another move), mostly has the chance to be evaluated differently. Think e.g. of a situation, where a node could be reached by a capturing move and also by some quiet moves from different ancestors. There are only very few nodes, having only one unique prior node.

Reinhard.