Author: Dann Corbit
Date: 16:48:14 01/14/04
Go up one level in this thread
On January 14, 2004 at 19:08:10, Uri Blass wrote: >On January 14, 2004 at 18:25:34, Tord Romstad wrote: > >>On January 14, 2004 at 16:56:20, Anthony Cozzie wrote: >> >>>On January 14, 2004 at 16:26:42, Ed Trice wrote: >>> >>>> >>>>> >>>>>Have you considered trying MTD(f) instead of PVS? I am not sure it is any more >>>>>efficient in practice, but it is easier to code, and has the additional benefit >>>>>of making you feel different, original, interesting, intelligent, handsome and >>>>>attractive. >>>>> >>>> >>>>Well Aske Plaat would love to hear that :) >>>> >>>>But doesn't MTD(f) trigger a great deal of researches? I remember trying that >>>>once and it bloated the tree. >>> >>>---- opinion mode on ---- >>> >>>MTD(f) has two big problems. >>> >>>1, you ponder the wrong move occasionally because your PVs are less accurate. >>>If you are pondering the wrong move 20% of the time that is equivalent to a 10% >>>time loss. >> >>This is not a *big* problem by any stretch of the imagination. It does indeed >>happen >>that the last few moves of the PV are wrong or missing, but I have *never* seen >>as obviously wrong move as the second move of the PV. This does of course not >>mean that it never happens, but it is clearly very rare. >> >>On the other hand, it *does* happen that the PV contains only one move, and >>there >>is no move to ponder at all. This happens maybe once every 5 or 10 games, but >>usually when the game is already won or lost. >> >>>2, MTD(f) is at its worst when the score is dropping. A fail high in MTD(F) is >>>much faster than a fail low (1 child node vs all child nodes). >> >>This is true. The average branching factor is clearly lower when the initial >>direction >>of the search is downward. >> >>>Unfortunately, >>>this is when you need your search the most: you are in trouble, and you need to >>>make exact moves to win/draw (you might already be lost, but thats just the way >>>it goes). >> >>Most of us extend the thinking time in such situations, and try to avoid making >>a >>move before the search fails high. >> >>By the way, there are a few things you could try to solve the problem you >>describe, >>although I haven't yet tried them. The main idea is to give up quickly if the >>search >>appers to fail low. The easiest thing to do is to just abort the search if the >>first move >>at the root fails low, and immediately start a new search with a lower search >>bound. >> >>It is certainly possible to find refinements to this idea, but as I said I >>haven't experimented >>with it yet. >> >>>I remember some Zappa-Gothmog games where Gothmog had been searching >>>8 ply, got in a tight spot, made a 6 ply search, played a huge blunder, and went >>>from -1 to -5 the next move. >> >>It is quite common that the search depth reached varies a lot from move to move >>in Gothmog (a difference of 3 or 4 plies is not unusual), but usually this is >>due to >>DFP rather than MTD(f). A sudden dramatic drop in search depth usually means >>that most of the DFP is disabled for some reason. >> >>And in general, if you want to knock MTD(f), you really need to base your >>conclusions on >>something more substantial than games against Gothmog, which undeniably is the >>slowest, weakest and most buggy MTD(f) engine known to man. > >I do not believe it. > >PostModernist also use MTD and I think that Gothmog is stronger than >PostModernist. > >I did not test Gothmog and my impression is based on results that I read that >suggest that Gothmog is at the same level of engines like Ktulu. I think you underestimate PostModernist. It may be stronger than SOS (often considered the strongest MTD(f) engine), and so it is surely one of the strongest MTD(f) engines around. Probably, though, Gothmog is stronger. Not enough data yet to be sure. 1: Quark 2.05b 41.0 / 68 XXXX 0=10 =0=0 1=1= 0111 0001 ==11 ==0= ==11 1111 0=11 101= 0101 =01= =0=1 1==0 1110 1111 2: Tao 5.4 39.5 / 68 1=01 XXXX 1110 001= 0101 1001 0001 ==== 101= 1==1 01=1 1010 01== 1011 =011 0=11 =1=0 11=1 3: Patzer 3.61 39.0 / 68 =1=1 0001 XXXX 000= =01= 001= 01== =1=1 001= =0=1 =0=1 0111 1=01 1=11 1101 110= 111= 110= 4: Amyan 1.59 38.0 / 68 0=0= 110= 111= XXXX 0001 0=== ==1= 10== =00= 101= 1001 1=== 100= 1011 =1=1 1==1 01== 11=1 5: Leila 0.53h 36.5 / 68 1000 1010 =10= 1110 XXXX 10== =010 01=0 11=0 1=10 0101 1==1 01== 11=1 1=11 =0== 01== 0001 6: WildCat 2.79 36.5 / 68 1110 0110 110= 1=== 01== XXXX 0=01 100= 10=0 00=0 ==11 1=== 1101 1001 =101 001= =1=0 1101 7: Comet B.60 35.5 / 68 ==00 1110 10== ==0= =101 1=10 XXXX =110 1=1= 1=01 1001 1001 1010 ==00 1=10 0011 0=01 1=0= 8: AnMon 5.22 35.5 / 68 ==1= ==== =0=0 01== 10=1 011= =001 XXXX 1010 =101 1=== 0011 0==1 100= =01= 0=11 =001 =11= 9: Delfi 4.3 35.0 / 68 ==00 010= 110= =11= 00=1 01=1 0=0= 0101 XXXX 0=0= 0=0= =10= 1100 1001 =011 1111 1==1 =101 10: Pepito 1.59 34.0 / 68 0000 0==0 =1=0 010= 0=01 11=1 0=10 =010 1=1= XXXX 1=1= 1010 1110 0110 =00= 0=01 10=1 111= 11: FrancescaMAD 0.0.9 33.5 / 68 1=00 10=0 =1=0 0110 1010 ==00 0110 0=== 1=1= 0=0= XXXX 0010 1000 1=11 10=0 11=1 1=10 ==11 12: Abrok 5.0 32.5 / 68 010= 0101 1000 0=== 0==0 0=== 0110 1100 =01= 0101 1101 XXXX 1101 1100 ==01 1=== 100= 1==0 13: PostModernist 1007 32.0 / 68 1010 10== 0=10 011= 10== 0010 0101 1==0 0011 0001 0111 0010 XXXX ==10 0010 110= 10=0 ==11 14: SOS 99.11.03 31.5 / 68 =10= 0100 0=00 0100 00=0 0110 ==11 011= 0110 1001 0=00 0011 ==01 XXXX 11=1 ==10 =1=1 0110 15: Phalanx XXII 30.5 / 68 =1=0 =100 0010 =0=0 0=00 =010 0=01 =10= =100 =11= 01=1 ==10 1101 00=0 XXXX ===0 0111 ==1= 16: Thinker 4.4h 30.0 / 68 0==1 1=00 001= 0==0 =1== 110= 1100 1=00 0000 1=10 00=0 0=== 001= ==01 ===1 XXXX 01=1 ==1= 17: Arasan 7.1 28.0 / 68 0001 =0=1 000= 10== 10== =0=1 1=10 =110 0==0 01=0 0=01 011= 01=1 =0=0 1000 10=0 XXXX 0100 18: KnightDreamer 3.2 23.5 / 68 0000 00=0 001= 00=0 1110 0010 0=1= =00= =010 000= ==00 0==1 ==00 1001 ==0= ==0= 1011 XXXX
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.