kbhearn wrote:There is generally no explicit randomness in the search. There's semi-random effects caused by timing, particularly in the case of an SMP search where some nodes might get searched under one timing and not under another due to threads getting preempted by the system at different points resulting in different nodes visited, different splits, even more different search order, different hash entries available at a given node, etc.
Threads imply parallel processing. Is that done nowadays? I assume so, and, if so, since threads can take different times to execute, I can easily see how this will create randomness akin (in my mind) to the same programs searching different parts of the same tree at the same time.
Also it's not 100% clear from Robert Hyatt's answer: do most engine searches nowadays truncate the tree for the sake of saving time in a blitz game, or not? And I'm not referring to Alpha–beta pruning, but simply deciding that certain moves in the chess tree will not even be considered beyond a minimum ply depth where all moves are considered; since, when playing at 5 seconds a move, I don't see how you can avoid doing this, since so little time is present, unless your program thinks on the opponent's time, which I believe is not done by most engines, in that they have a parameter to turn this think on opponent feature off.
kbhearn wrote: The most common case for the weaker engine winning though is where a line is bad for the stronger engine but neither the strong nor weaker engine see it and the weaker engine just falls into 'oh, i'm winning!', the stronger engine likely even seeing it first but not being able to do anything about it by then.
Yes, thank you, I understood this 'ignorance is bliss' argument, and in fact when I play stronger human players, I sometimes win the same way (they see too much for their own good, while I stick to some straightforward tactical theme and sometimes get lucky and beat them).
Yes, but I wish somebody would collect these oddities so we could examine them. I once saw a thread where each strong engine was given 24 hours a move on fast hardware, and though the game ended in a draw, it was quite fascinating (as I recall it involved a queen for two rooks trade, which is, next to a queen for three minor pieces trade, interesting to watch).kbhearn wrote: The less common case of seeing more is bad should be very uncommon these days but would involve a long line where a large unavoidable advantage occurs at the cost of a disadvantage that takes an even longer line to see how to defend it and have the disadvantage you gave up become the relevant issue. At current search depths the chance that the root evaluation relies on such leaves is... rare.
JayRod