Moderators: hgm, Rebel, chrisw
hey oviOvyron wrote: ↑Tue Nov 26, 2019 4:48 amYes, and people take advantage of it to win correspondence chess games.Spliffjiffer wrote: ↑Tue Nov 26, 2019 4:45 am...ok its very often the case that the PV is NOT accurate, weather in SF nor its derivatives nor in any engine ??
...you respond so fast that even my correction from "weather" to "wether" was not fast enough, sure i give u right Don't worry, Spliffjiffer, because you will have to correct it again (correctly this time), from "wether" to "whether".Spliffjiffer wrote: ↑Tue Nov 26, 2019 5:00 am hey ovi
omg, a castrated sheep...no, that was not i wanted to tell !!....whether, whether, whether, ok...some exercise, thank you very much jp !...ill try to learn itjp wrote: ↑Tue Nov 26, 2019 1:13 pmDon't worry, Spliffjiffer, because you will have to correct it again (correctly this time), from "wether" to "whether".Spliffjiffer wrote: ↑Tue Nov 26, 2019 5:00 am hey ovi
That must refer to NN engines running on normal people's hardware.
Exactly. So I still don't get why people would sit on the root position since only getting 1 ply doesn't change with more time. When the PV gets inaccurate I call it the "tail of the variation." If the first move on the PV is a mistake (because it changes the theoretical result of best play) then it's only a tail with no headDann Corbit wrote: ↑Tue Nov 26, 2019 4:57 am A pv is only accurate to the point where the opponent engine or player makes a different move.
So you get 1 ply, and maybe more.
I like this picture:Ovyron wrote: ↑Wed Sep 18, 2019 4:40 pm(using code tags so this wall of text doesn't shoot my message's length to oblivion)Code: Select all
The "weakness" of A/B engines is that their PVs are garbage. Seriously, I have tried every single analysis method I could think about with them, the most successful ones by far were the ones where I put in my analysis tree the move that the engine played (the 1st move in its PV), and threw away the rest, which was never useful (if you use MultiPV, you take note of all the 1st moves on the many PVs, and throw the rest away, or you can exclude the best move from analysis, see what's the second best, and throw the rest of its PV away). The problem with engines is that, for this PV that they show there's a "weakest link", that is, a move that if they have the position in their face, they wouldn't play at all. Never. If you can find this move, you can refute their line. What happens is that on a given PV, only the main move is analyzed to this depth, the -next move for this side- that appears will have been analyzed to 2 depths less, and the next one to 4 depths less... For an analysis to Depth 70, from move 11 (which is the case here), the engine will show a line like this: 11.A a 12.B b 13.C c 14.D d 15.E e 16.F f 17.G g 18.H h 19.I i 20.J j 21.K k 22.L l 23.M m 24.N n 25.O o 26.P p 27.Q q 28.R r 29.S s 30.T t 31.U u ... Now, due to reductions and extensions, what follows isn't -precise-, but it is -accurate-, only A is at depth 70, you can see how the move quality of the rest of the moves goes downhill like this: A/70 Depth B/68 Depth C/66 Depth D/64 Depth E/62 Depth F/60 Depth G/58 Depth H/56 Depth I/54 Depth J/52 Depth K/50 Depth L/48 Depth M/46 Depth O/44 Depth P/42 Depth Q/40 Depth R/38 Depth S/36 Depth T/34 Depth U/32 Depth That is, to get a better PV than the one you have, you don't need to wait till depth 71, all you need to do is jumping into any of these positions, and reach what I call "relative depth +1". Say, you can jump to 24...n and reach depth 45 on there, or to move 27...q and reach depth 39 on there. But what you want is to find the weakest link, where is it? Who knows, but you'll notice the closer you get to the root, the longer it takes to reach "relative depth +1." So you can follow the next procedure: 1.Jump to move 30...t and reach depth 33. IF score == 0.00 THEN go to 2. ELSE score >= 0.00 you refuted the line! Congrats! 2.Jump to move 29...s and reach depth 35. IF score == 0.00 THEN go to 3. ELSE score >= 0.00 you refuted the line! Congrats! 3.Jump to move 28...r and reach depth 37. IF score == 0.00 THEN go to 4. ELSE score >= 0.00 you refuted the line! Congrats! ... If you do it all the way up to 11...a at Depth 69 and you got 0.00 all along, then I concede, Stockfish's line is solid, and it is 0.00, and I couldn't have refuted it, blah blah blah. Otherwise, you'll see that there was some garbage in the PV, but that Stockfish couldn't see at this relative depth, because it was outside its horizon. You can make the horizon larger in a much faster time than it takes Stockfish to reach Depth 71. Heck, it takes Stockfish so long to get to the next depth, that you probably can do "relative depth +2" or "relative depth +3" and get ahead of Stockfish's depth when it is done! What do you do when you've refuted this line? Well, you know from the root that it's 0.00 because if this falls Stockfish has another 0.00 line somewhere. You can explore the variations until you find it, and then you can use the same trick to refute this line to a >0.00 score. Once this is done, you can find another one and find the same. Eventually, Stockfish will run out of 0.00 lines, so you'll be able to have a root score that is >0.00. You'll have a score of >0.00 even though Stockfish at Depth 71 or Depth 72 will continue to show 0.00. Why is this? Because the moves that take it down are still outside of its horizon, and there's moves it will just not see unless the relevant position is on its face (sometimes it will not even see it on its face, you need to use MultiPV on there, or use a different engine that sees the move so you feed it and it sees the truth.) And now that the root position has a >0.00 score in your tree, you can do this all over again, but shotgun all the variations with this score, which you can call S, and after your root position has score S, you can do this all over again to make the root position >S on your tree, and so on. What you'll find out if that this is a much more productive endeavor than letting the engine unassisted with Infinite Analysis for hours. SO MUCH that I could easily let my engine analyzing some position unassisted while I'm asleep, having some high depth PV when I wake up, but I don't do it, because it's a waste of time, because only the first move of the PV is useful and the rest is garbage, because you're much better off building a consistent line that has the same relative depth for the moves like this: A/26 Depth B/26 Depth C/26 Depth D/26 Depth E/26 Depth F/26 Depth G/26 Depth H/26 Depth I/26 Depth J/26 Depth K/26 Depth L/26 Depth M/26 Depth O/26 Depth P/26 Depth Q/26 Depth R/26 Depth S/26 Depth T/26 Depth U/26 Depth Because Depth 70 continues V/30 W/28 X/26 Y/24 Z/22 - And you'll have Y/26 Z/26 - You'll be outsearching Depth 70 by 4 depths and they'll continue to go up after move Z, where Depth 70 will be relative Depth 20, and you can refute it in a second.
The point is that the evaluation is -1515, meaning that Stockfish recognizes that the position is lost for Black. I should have simply omitted the PV.