On 5/2/2024 2:20 PM, Philippe Michel wrote:
On Wed, May 01, 2024 at 05:31:02PM -0600, Murat K wrote:
[...] Could this kind of bizarre games have been happening in rollouts for all those years without anyone noticing? And if so, what would the implication of that would be??
I suppose they do happen, but very rarely.
How do you know to say "rarely"?
With the cube in play they will often be stopped at some point by a double/pass and in the case of
This is what I would have expected and expressed also. In those cases there would be nothing unusual to talk about anyway. In the example I encountered, the cube was never turned in, who knows how many, hundreds of turns in a single game. I suppose the opposite could also happen, with the cube going insanely high because of double/takes back and forth, as the game goes in circles and the same positions keep repeating over and over again until the bot somehow breaks out of the loop.
a cubeless game the result is always between -3 and +3 points, in one trial among many.
Score would matter only in my experiments, (and if we ignore win rate in proportion to the error rate, at that). Talking about rollouts, total error rates accumulated from all those hundreds of weird moves may get high enough to effect the results, no? I don't know rollouts all that much. You guys would know this better.
A rollout of a position like the one you shown would probably be suspect since every game would start with wild, poorly evaluated positions, but a rollout starting from a relatively normal position should be fine even if one or a few trials are unusually long.
The first position I posted was the first one I could catch long into the game, after probably many hundred weird positions, increasingly getting worse. The sequence probably started with a "relatively normal" position, as you said. Unusual length of games/trials is not the real issue. Depending on why the trials are done, possibly astronomical cube values and/or error rates may be problems. In either case, I don't think neither you nor anyone else can know enough to say that they are "few" or "rare". MK