🚀 go-pugleaf

>>I've noticed the following, and can find no explanation for this:

>>Set the depth for gnuchess evaluations to, say, 6 ply.
>>Now set post and go to evaluate a position.
>>This should yield some score and a best move.

>>Now in the new position, set the depth to one fewer ply, 5.
>>Again, press "go" and watch for a score.
>>Compare the new score with the old one.

>>They're different!   Why?

>There are some possibilities:

>* GNU Chess chose parameters for the evaluation function depending
>  on the current position.

>* GNU Chess add a random number as part of the evaluation function.
>  (This can be turned off)

Thank you for the information.  Right after I posted I realized it
might be due in part to the random number generator (who wants to play
the same game every time?). Your first point about the position dependant
evaluation function gives me pause for thought...  I never considered
that as a possibility.  I thought the best information would come from
the deepest nodes, but I guess the initial position is useful to direct
the search.

>* The evaluation function depend at least partly on the Alpha-Beta
>  interval. If a fast approximation is outside the interval by a
>  large enough margin, the slower evaluation is not used.

Your fourth point made me realize that more than one eval function may be
in use.  That is reasonable to explain this since it based on the a-b interval.
Great!

>* The order of the moves depend on the killer move heuristic as well
>  as the history heuristic. The order the game tree is searched
>  depend on what has been searched before.

The fourth point only concernes the order the tree search happens.  I guess the
killer and history heuristics depend on the initial position, but how would
that influence the result score for fixed depth?  That should only affect the
extensions.

OK I get it.  The depth for extensions is set separately...  <blush>

>>This doesn't always happen.  The scores (accounting for
>>who goes next times minus one) are usually within about 1/10
>>of a pawn.  But shouldn't they be exactly the same?

>I think my first reason is the most important one.

>>By the way, book is off, transposition table is off.

>You could turn random off too.

Thanks for your explanations! I though no one was reading this group!!

Now that I have some idea why Gnuchess does this.  *Why* should it give a
different eval under these conditions.  Shouldn't the overall result be the
same at 5 ply for the best move as for 6 ply?


Mike