Author: Ulysses Omycron
Date: 18:31:00 01/02/05
Go up one level in this thread
On January 02, 2005 at 20:53:33, Uri Blass wrote: >On January 02, 2005 at 17:57:16, Ulysses Omycron wrote: > >>On January 02, 2005 at 17:19:52, Peter Berger wrote: >> >> >>No, it doesn't, it seems it actually scored all the options, > >It scored all the option in the way that peter described(first searching the >best option and later searching the best option when the best is ignored). No, if it did so, the search would take twice the time if you search 2 best moves, threce if you searched 3, and then go up exponentially since it needs to do all the work again ignoring the best move. It doesn't take more time than normal search. In a very hard position I still reach ply 15 with all options scored within one minute in my AMD 900 (Ok, so I may be lying, but this discussion makes no sense without a real example, give me a position to analyse and you may be right). >>the output to show all the moves, you may get (For example) that axb4 scores >>5.34, a suicide move, but the scores would be used to sort the next iteration, >>so if bxc3 scores 9.80 (The worst available move) it will be searched last in >>the next iteration. (At least that's what REBEL 12.00.02 for MS-DOS seems to do >>with the worses moves, but stills searches the most interesting [After the best >>move] first). > >programs often do not search last the worst move in one best mode because they >do not know which move is the worst. They know if ChessBase tells it to score all moves, so the one with the lowest score is the worst. Have you seen the worst move turning out to be the best? I neither, but I've seen the 3rd worst move turning out to be the best, when you are deep position anylising with the total time option, sorting the moves in the best may save time, but may be risky (That's why I don't use the total time option... Err... I think in some point I contradicted myself). >> >>If you are right, then what ChessBase does is actually trickign the engine, >>making it think that it already made the notasgood move, so it gets an score >>because the best moves after it are searched and so and so. This is what >>CM10thEd seems to do in Mentor Lines. > >Chessbase does nothing except telling the engine how many moves to search. > >Part of the programmers support that option by using the algorithm that peter >suggest. > >Note that they may do things faster by not repeating the same iteration 3 times >for 3 best move. > >The idea is that after you have 3 moves at depth 7 and starts depth 8 you get an >exact score for all the 3 moves that you evaluated as best at depth 7 and for >every additional move you search if it is better than the third best move and >you calculate exact score only in case that it is better. That's why I run DPA several times at several conditions. Running a big DPA one time with lots of time and very deep is going to give wrong results, that's where the human intelligence comes into effect: If you see that the engine is doing senseless moves with a very high score and ignore human promising moves you imput them, and do DPA for that move, the move after it (To get a more exact score) and the move before it (To be sure that nothing is left behind), do this several times deeper and deeper and you eventualy solve the position. DPA by itself is worthless, you need to know how to use it. Give me a position and I will give you the best moves available for one side (If I fail, you're right). >>Well, if Qc7 is really the best move, eventually all the other moves after being >>DPAed will drop their scores, and Qc7 will raise its score till it's the best >>move, so you don't care actually for what move the engine is going to do, but >>for what is the best move after DPAing it very deep. > >You have illusion that Deep position analysis can solve it. > >It can often not solve it because deep position analysis may miss the next move >after Qc7 2.Rc1. It eventually will too. > >The best move after 1....Qc7 2.Rc1 can be only number 10 in the list of moves. > As I said, the human may imput those moves. And don't say "If the human already knows the best move then why to analyse with the computer?" as I'll answer by: "...Qc7 2. Rc1 what? Give me a real position to DPA without giving me the answer to it, and I will solve it at some level". >Deep position analysis is practically not better for playing strength than >normal mode of the engine(otherwise programmers could tell the engine to use >deep position analysis in games automatically). Have you compared the results? You might be surprised (Using it correctly, not just 1 DPA). Programers should tell the ngine to use DPA in games as the moves are more reliable (But it takes more time, I'm not talking about using DPA for playing [Unless from the starting postion if you already have solved it at some level], but to know what is the best move available in a position after a Deep Postion Analysis). >>fact that if you never do e4 (For example) you don't need to know the Petroff >>Defence, Ruy Lopez, 2, 3 and 4 nights variations, you just need to know what >>it's the best defence against it, if it turns out that the french e4 e6 d4 d5 is >>the best, then you have succefully prunned way more than 1000000000 moves and >>1000000 variations already, since you are doing d4, you don't need to have them >>on the book." > >All engines do exactly this by the alphabeta algorithm in their normal mode >without deep position analysis. But they fail. >If they want to play 1.d4 and 1.e4 e6 is worse than 1.d4 for white based on >their analysis they will not analyze other the petroff defence when they analyze >1.e4 But it may turn out that d4 e6 turns out to be worse than e4 e6, and after another DPA the petroff is more dangerous than e4 e6 making it to return to d4. DPA is pretty dynamic, just don't use it once and say it's bad. But these are only examples, we need some action of real analysis! Happy Hunting
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.