Author: Robert Hyatt
Date: 05:57:27 04/03/00
Go up one level in this thread
On April 02, 2000 at 22:39:06, Tina Long wrote: >On April 02, 2000 at 11:03:56, Robert Hyatt wrote: > >>On April 01, 2000 at 23:17:11, Tina Long wrote: >> >>>Hi guys, >>>This thread has been interesting & generally agreed with previous stuff I'd >>>read. >>> >>>So. >>>Analysing 3 best moves, over 24hours, slows the program by a factor of 3 - 6, >>>losing 1 or 2 ply depth. >>> >>>Each extra Ply has a 15% of picking a different move, that the software >>>calculates is better. >>> >>>But. >>>What about the Actual ply depths searched on the 2nd & 3rd preffered moves? >>>If only the best move is shown, no's 2 & 3 share with no's 4... about 50% of the >>>search time. >>> >>>If best three is shown, no's 2 & 3 are receiving 6 - 10 X more thinking time. >>>1st preferred move has gone from 50% to 25% of thinking time (25% for "rest") >>>or 1st preferred move has gone from 50% to 17% of thinking time (50% for "rest") >>> >>>preferences 2 & 3 are being searched an extra 1 - 2 - 3 ply deeper over 24hrs. >>>Their "score" is more acurate than if only 1 best is shown. >> >>I don't follow "being searched an extra ply and their score is more accurate >>than if ..." > >By looking at 3-best instead of 1-best, the 2nd & 3rd best moves are examined >about 6 X longer in time. In 24 hours they are searched for about 4 hours each, >instead of about 40 minutes each. All I'm saying is that there must be more >accuracy in the "score" for those moves. If the 6X time doesn't result in an >extra ply or 2 of depth, then the actual amount of computing at the reached >depth must achieve something, or else be redundant. > That isn't how alpha/beta works. In the normal case, for the second move, we only have to prove it is worse than the best. In 2-best mode, we can't stop when we prove it worse, we have to spend all the extra time to get the true score. We don't search it any more deeply. We don't learn anything new, since we already knew it wasn't as good as the first move, we just spend more time to get the true value rather than using the traditional alpha/ beta cutoff to stop the search early and save time. Note that alpha/beta search and minimax search are provably identical in the result they produce. But that alpha/beta trees are _much_ smaller to get that same result. >> >>Here is the main point: If the best move looks best thru 19 plies, but is >>going to look horrible at depth 20, the 3-best display may ensure that you >>only get to 19, where you would get to 20 normally. So the normal search >>will see that the best move is horrible and find an alternative if one is >>available. The 3-best search will only get to depth 19 and think that the >>original best move is the one to play. > >Granted, something may be missed by not moving from the 19th to 20th ply. But >how much is being missed in the examination of 2nd & 3rd best moves when only >1st best is shown. > Not a thing. See the original paper on alpha/beta, or the follow-up written by Knuth and Moore. What you are 'thinking' is happening is _not_ what alpha/ beta is actually doing. >I am suggesting that "If the "score" of the best 3 moves is not too dissimilar" >increasing the analysis time from about 40min to 4 hours for 2nd & 3rd best May >find one of them to be actually Best. This is the Benefit to weigh against the >Cost of the 1st best move being analysed to 19 ply rather than 20 ply. This is the same thing as adding 2+2, but then going over and over the result, checking yourself, for the next hour. Is your answer any more accurate? Or did you just spend a lot of time? If the second move was searched deeper, that would be different. But _all_ moves are searched to the same base depth. > >> >> >>> >>>So. >>>If the "score" of the best 3 moves is not too dissimilar, then "forcing" the >>>program to look at choices 2 & 3 for longer could be a good thing. >>> >>>Eg. >>>guessed ply depths "well" searched for the first four choices over 24 hrs. >>> >>>Show 1 best: 1-19 2-15 3-15 4-15 >>>Show 3 best: 1-17 2-17 3-17 4-13 >>> >>>Is the "danger" of reducing the favoured move ply from 19 to 17 greater than the >>>"benefit" of increasing the 2nd & 3rd ply from 15 to 17? >> >>I don't follow how the second/third moves are increased. They are _all_ >>searched to the same depth each iteration, even in normal mode where only the >>best move and score is displayed. But _all_ moves are searched to this depth. >> >When only the best move is displayed, 50% of the time is spent examining that >best move. All other moves share the other 50% of time. If they are all >examined to the same ply depth, then many more "somethings" must be being >examined in the best move compared to the 2nd best move. > >Is there any benefit in examining more "somethings" (for 6X the amount of time) >in the 2nd & 3rd best move, at the cost of an extra ply for the best & all other >moves. Nothing useful is found. Here is an example: there are three holes in the wall. Your task is to stick your hand in each hole for one minute, and at the end, report which was the most pleasant (or least unpleasant). You stick your hand in hole 1, and for 60 seconds you get nothing but warm water run over it. You stick your hand into hole 2, and immediately get stuck by a pin. Do you wait longer to see if you get stuck by a knife? burned? Or do you stop _now_ since you know the pin stick is worse than 1 minute of warm water? I stop. On to hole number three where you are greeted by salt water that is just below the normal freezing point of water. Do you exit immediately, knowing the warm water is more pleasant, or do you wait for the salt water to stop and the chain saw to start? You can spend just over 1 minute, or a total of three minutes. If your goal is to rank the three holes from best to worst, you have to take 3 minutes so that you know _everything_ that happens. But if you only have to pick the 'best hole', just over 1 minute is more than enough. > >I realise I'm not being clear, but all that extra time being spent on 2nd & 3rd >best, must be doing something for the accuracy of their analysis. > >Thanks >Tina Long > > > > >>>Tina Long
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.