Author: Thorsten Czub
Date: 04:31:32 08/19/04
Go up one level in this thread
On August 18, 2004 at 23:42:50, Albert Silver wrote: >:-) Actually, I'm not saying these are the best possible, just that they've >yielded good results so far, and more importantly: good positions. yes. same with my style. we don't really know for sure until we have played enough games. i have your style with cleared hash table change in an eng-eng match against S8 on the pc in the moment. and i will see how good it scores. then i do test my style. >One of the things I was unhappy with when watching Pro Deo play Junior 8 was >that even when it won, it seemed a bit lucky at times. Of course, it had its >solid wins, but many, no, most of the time Junior 8 had the initiative and not >merely an illusory one. I'll run a new battery of tests again tonight. No doubt >time and tweaking will improve on the settings even more. this is in the default style with S8 and F8 the same as you describe. it has IMO to do with the fact that ed is not giving the default style high values. this way pro deo is not taking the initiative although it could. this changes in the moment you play arround with some switches. only you have to find a setting that works for all. >Maybe. I got this from the Rebel 10.5 page. It illustrated a couple of positions >where knowledge helped it spot a key move 2 plies earlier when set to 200 as >opposed to the default 100. But as he explains in the part on the lazy eval, >many times the extra knowledge is unnecessary, hence the whole concept of the >lazy eval. Plus it slows the search down, so the tricky part (especially having >no idea what the program knows or doesn't) is in finding a balance where the >slowdown is worth it. here i cannot disagree. only that i have a different point of view. IMO the search is completely unimportant, for me the evaluation is the boss, not the search. the search is helpful, but the score is much more important IMO. when looking into your style it seems you trust on Ed's % comments in the REM part and you try to increase the search strength by using those switches Ed claims to get + x %. I doubt that using them at all works. but we can try out different settings and styles of course. IMO the problem is that as long as you don't have EVIDENCE that the % is true, you cannot trust in the remarks. my golden rule is: never trust a programmer when he says : + xy %. this is almost never true. I have no idea how this works. But it seems a very basic law in computerchess. when a programmer has a result, it is often nothing else than statistical fog or white noise. or the Pc is broken or he tested a different program / exe or or or. whenever i get a program that is new, i first try the default. and if the default is not giving the % the programmer or the company says, i am changing the parameters to see if the program is somehow useful. > I'll continue testing with it as is, and see what it >yields. I use the Nunn sets because I can also see exactly where and how it is >changing its behavior and how it is developing the position. It goes without >saying I'm not tweaking it for a specific move or moves. i see. but you have to knock off the learning for this (in the style). in your style the learning was ON: [Book Learner = aggressive] off|passive|moderate|strong|aggressive [Position Learner = on] off|on BTW: i do not test this way because i would be bored by always the same NUNN/NONE openings. IMO the Nunn-Positions are NOT good test-positions to start from. John Nunn might be a good chess player, but IMO his positions are not very good chosen. but the people use those positions because someone once started with them. >Borrow away! You're not stealing. I just want to see it play its best game, >which I'm not convinced it is yet (with all due respect to Ed). i have the same motivation. >>[King Safety = 160] * >>[King_Safety = 180] * King Safety new style (!) > >>[Traditional Isolated Pawns = 115] * weight isolated pawns >>[Progressive Isolated Pawns = 110] * weight isolated pawns (new style) > >Isn't the new style supposed to replace the older one? If so, then shouldn't one >simply remove the older reference? That's the way I understood it, which is why >you'll only see one reference in mine BTW. I do understand the logic in your sentence. it should normally be replace the old knowledge. on the other hand ed's explanation / comment he gave about this was not 100 SURE about the superiority of the new method. so why not using both :-)) maybe this keeps you on the save side of the bridge. >>[Own Queen = 1.00] * Don't exchange queens when down (!) >>[Pruning = MISC_13] * Bishop Pair (extra bonus) (!) > >I didn't add these on purpose since I want to see enough cases where they seem >necessary. Well, at least one clear one. Also, I'm a bit afraid of sprucing up >the bishop pair advantage too much, simply because it may decide to hold on to >that bishop pair incorrectly. yes. it depends how YOU see it. imo 2 bishops in an attack situation with certain kind of pawns are superior. in other situations 2 knights are better. difficult to find a setting that works for all. >>[Pawn Pressure midg = 150] * weight pawn pressure middlegame >>[Pawn Pressure end0 = 200] * weight pawn pressure early endgame >>[Pawn Pressure end1 = 200] * weight pawn pressure endgame >>[Pawn Pressure end2 = 100] * weight pawn pressure simple endgame > >I only put the first two above, since I'm not sure what is meant by the pressure >in the endgame. i guess to push passed pawns for the promotion. >>[Pawn Formation = 105] * 100 >>[Weak Pawns = 110] * 100 >>[Passed Pawn King Tropism = 105] * king tropism to own passed pawns > >I may add these. As I said, it's early for me to start changing. This is only my >3rd personality to be honest. if i would count the amount of personalities i did with rebel/pro deo, i could have had thousands. >>[Chess Knowledge = 235] * > >As discussed above. yes. i like it higher :-)) >>[Attractiveness = 110] * >>[Attacking = 110] * >>[Strength of Play = 100] * >>[Draw Contempt Factor = 0.00] * >>[Selective Search = 115] * >This one I asked Ed about but got no answer yet. The added value to the >selective Search is not clear to me. At first I was inclined to think it was >like the King or Hiarcs, but after reading his description in the parameters, I >see it as cutting out or increasing the use of Null-move compared to using it >together with his new idea. Note that he says that reducing it to zero >(Selective Search = 0) would make it a full null-move program. i tried it out. 100 was IMO not deep enough. 130 too much. 115 seems ok. but this maybe depends on the other search stuff you have switched on. >>[Pruning ? NONE] * >>[Search Technique = NULLMOVE] * >>[Engine Learner = off] * off|on >>[Book Learner = aggressive] * off|passive|moderate|strong|aggressive >>[Position Learner = on] * off|on >>[Extended Book Learner = read&write]* off|read|write|read&write >>[Keep Main Line ? 10] * minimum margin new mainline (?) >>[Pruning ? MISC_50] * Limit Reductions table driven (=) >>[Checks Depth ? 1] * limit QS checks to 1 ply (+9%) (!) > >Althoug I believe this works, I'd like to hear an explanation on it. Just >curious. yes. i have no idea what check depths or keep main line have an effect. i cannot see that those switches do work. also i am very very suspicious with those "reductions". IMO they often delete important lines. >>[Recaptures = 2.00] * 1.50|2.00|2.80|3.00 > >I'm not going to touch this as I have no idea what it does. i have not found out any big advantage or disadvantage when using different values for this function. it seems it is neither weakening nor making the program stronger no matter which result you input. >Interesting game, but I'll be honest I thought it a bit chaotic. Chaotic ? IMO the steady increase of the evaluation is not chaotic ! its the opposite. this was no chaotic game, it was a clear win by pro deo. shredder had no clue to win. A nice win for >sure, but I am interested in seeing how it handles slightly more normal >positions and how it develops them. Still, do share the neat attacks or >combinations. :-) > > Albert it all begins with NORMAL positions. then pro deo has to take initiative and then it escalates in an attack with win. that should be the NORMAL way you beat a chess program. important is IMO that your evaluation function is capable to recognize weaknesses in the opponents positions. therefore you need functions that influence the overall score. with no such functions, you cannot hope to recognize weaknesses. you would MAYBE find those weaknesses if you search deeper than S8, but this is a tough job to do :-)) btw: in the moment Pro deo is losing +0 =1 -2 against S8 with your style and cleared hash ON.
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.