syzygy wrote: ↑Sun Aug 29, 2021 2:59 amA horizon effect would mean that the search manages to push a tactical trick behind the search horizon, which seems rather unlikely here.
This may not be the standard/consensus definition among chess programmers, but I would say the horizon effect is when anything is pushed beyond the search horizon. In my view it need not be a tactical trick, it may simply be recognition of some positional feature, as well.
It does seem unlikely that any modernish A/B engine with quiescence search and tactics-oriented extension heuristics could be caught flat-footed by anything but the deepest tactics, but I could certainly see even Stockfish 14 failing to notice a relevant positional feature until it's much too late. There are still test positions and puzzles that Stockfish needs a helping hand to solve.
dkappe wrote: ↑Thu Aug 26, 2021 9:40 pm
With the Stockfish nets getting ever bigger, a small but vocal minority has voiced concern that a drop in NPS could somehow harm its fitness as a chess engine. Maybe it is vulnerable to being “outsearched” by a strong, nimble competitor.
I decided to run a test with SF10 against its successors to see if the slower ones do suffer from some sort of horizon effect and fall off a cliff. The conclusion? Nope. The bigger the net, the more resistant the SF version seems to be to being outsearched.
