New testing thread

[Uri Blass]

I clear the killers and I do not think that it is very important.
I am going to get new killers very quickly in the first plies of the search.

Uri

Suppose you do the logical thing, however, and if you complete a 20 ply search, and start pondering, you start at ply=19 and not ply=1? Why? you made the first move in the PV, you are pondering the second, so you still have the result of an 18 ply search already computed in the remainder of the PV. So starting at 19 avoids repeated work. And now, suddenly, a good killer at ply=2 becomes very significant.

You are right but practically I start pondering at ply 1(I agree that starting pondering at ply=19 without clearing the killers may be better for playing strength).

Uri

[hgm]

I agree tha sn=n is a problem. And in Crafty, st=n is also a problem because if you set a specific time, I can't override it on fail lows. But it would take a lot of thought to come up with a sn equivalent that would be fair to all

Well, using a node budget as you would be using a time budget goes a long way in the right direction. This is why I chose to implement node-based time controls that way, in WinBoard, and not like the sn=n. You can still acheive the same by using nps=N and st=T in combination with each other, but you can do much more.

Playing by nodes can never be completely fair, as 'nodes' are not a universally defined concept, and not all engines work exactly the same, and even if they do, they might count nodes differently. If the nps rate is very variable, the reason for it is usually pretty obvious as well, and engine authors might be able to correct for it (e.g. count a tablebase probe as 10 nodes, if that is what it takes to get their reported nps approximately the same in the end-game as in the middle-game).

If there are any ideas on how to do something better, I am open to suggestions.

[Tony]

Maybe I missed it but wouldn't most be solved if the checkrate goes up when the nodelimit is almost reached ?

ie if 1M nodes left, check every 100K, if 300K left check every 50K etc. That way, one should be able to come to within 0.1 %of the nodelimit without big slowdowns.

Tony

[OliverBr]

In any case they are doing more than you think. OliThink was recently removed from the CCRL for no other reasons than that it did inexpliquable bad moves in the end-game, occasionally. No crashes, no illegal moves, no refusal of moves, no time forfeits, no sign of any trouble other than that the testers did not like the moves it played.

Ups. This is the first time I hear about that!
Which version of OiThink? What rule has been broken by OliThink in this case?

Is there some source where I can see a discussion/games about the incidents?