Re: AlphaZero No Castling Chess
Posted: Sat Dec 07, 2019 7:39 pm
Actually, after I posted, I thought of the very same idea! Yes, you did it right, but 25 is way too small a value for this situation. I was thinking of using 75 or 100, go with 75 if you want to be conservative. That would roughly simulate how I would think about the situation as a human.Laskos wrote: ↑Sat Dec 07, 2019 7:22 pmAh, forgot to specify:lkaufman wrote: ↑Sat Dec 07, 2019 6:54 pmWow, that's a surprise, rather different result from what S. Pohl got, but he used "normal" openings so I think your result is more likely correct. I don't think you are "missing something", it's just that humans much prefer to attack than to try to hold a bad position, and so nearly every GM would rather play White in this scenario, and most likely White would score well over 50%. Engines can't replicate human behavior perfectly yet, and so such simulations are only a rough guide to what would happen with human play. But your result is quite significant, it may mean that this variant is ideal for correspondence or engine vs engine matches, and perhaps even with human OTB play White's advantage would be within acceptable bounds given that the players always play two game sets. But then tie scores become likely again!Laskos wrote: ↑Sat Dec 07, 2019 6:36 pmI built an opening suite with Komodo on 4 threads and Variety option, almost 500 different 5-movers (the unique starting position is set to no Black castling):lkaufman wrote: ↑Fri Dec 06, 2019 4:25 pmI was involved with this idea too at the time, along with variants of it such as Black can only castle long, or white can only castle short while Black can only castle long. The pure version you tested seems to favor White too much even with Black winning draws. A more aesthetic version that might be more balanced (perhaps a bit in Black's favor between engines) would be that White can castle normally, but neither player can castle on the same side as his opponent has done. But this requires new programming, not so trivial to test. My subjective opinion as a GM is that the pure version you tested would be seen as too favorable for White (even with the Armageddon rule) for human play.Laskos wrote: ↑Fri Dec 06, 2019 1:46 pmAh, sorry, I didn't know. So, I re-discovered this morning meddling about handicaps your proposal. Seeing your results, it seems a viable proposal, if Kramnik thinks no castling at all is a viable proposal. I am not sure how many top GMs share Kramnik's view.pohl4711 wrote: ↑Fri Dec 06, 2019 12:13 pmYou have an even simpler modification? That is my NBC-Armageddon openings idea. Not yours. I released it in August 2019, with 4 openings-sets. And announced it here on talkchess.
From my website:
Level 2: NBC (= No Black Castling): White can castle to both sides, black is not allowed to castle. Line: 1. Na3 Nh6 2. Nb1 Rg8 3. Na3 Rh8 4. Nb1 Ng8 5. Nc3 Na6 6. Nb1 Rb8 7. Na3 Ra8 8. Nb1 Nb8
Level 2 (NBC) testing:
White Wins: 296 (59.2 %), Black Wins: 204 (40.8 %), Draws: 0 (0.0 %) White Score: 59.2 %, Black Score: 40.8 %
1 SF 190728 (half time) : 3477 500 (+304,= 0,-196), 60.8 %
2 Komodo 13.01 : 3400 500 (+196,= 0,-304), 39.2 % (Elo-spreading: 77 Elo)
White Wins: 284 (56.8 %), Black Wins: 216 (43.2 %), Draws: 0 (0.0 %), White Score: 56.8 %, Black Score: 43.2 %
1 SF 190728 (half time) : 3451 500 (+286,= 0,-214), 57.2 %
2 Komodo 13.01 : 3400 500 (+214,= 0,-286), 42.8 % (Elo-spreading: 51 Elo)
White Wins: 287 (57.4 %), Black Wins: 213 (42.6 %), Draws: 0 (0.0 %)m, White Score: 57.4 %, Black Score: 42.6 %
1 SF 190728 (half time) : 3490 500 (+313,= 0,-187), 62.6 %
2 Komodo 13.01 : 3400 500 (+187,= 0,-313), 37.4 % (Elo-spreading: 90 Elo)
White Wins: 277 (55.4 %), Black Wins: 223 (44.6 %), Draws: 0 (0.0 %), White Score: 55.4 %, Black Score: 44.6 %
1 SF 190728 (half time) : 3473 500 (+301,= 0,-199), 60.2 %
2 Komodo 13.01 : 3400 500 (+199,= 0,-301), 39.8 % (Elo-spreading: 73 Elo)
Thanks for opening suites too! These are from normal human and computer games? I guess the openings will be somewhat different if one side doesn't castle, maybe I will build with Lc0 and Komodo some short opening suites which are played for this variant.
http://s000.tinyupload.com/?file_id=235 ... 1942827279
The pure No Black Castling seems balanced with that Armageddon scoring in Komodo self-matches at different time controls from this suite, for example:
White wins: 101/200
Black Wins: 5/200
White wins: 20/40
Black Wins: 2/40
But if you feel that it is too much an advantage for White, maybe I am missing something with these engine matches (Lc0 also sees it as pretty balanced).
Well, one point you may have missed is that if the engines were aware that draws counted as Black wins and modified to play accordingly, White would surely have won at least a few of the drawn games, so I would say that the result does show some White advantage, but perhaps not too much. Humans would avoid anything looking even somewhat drawish as White, preferring an unclear attack to a small endgame plus.
Contempt = 25
White Contempt = True
Does that make sense to you? I am not sure how Black behaves in this case, maybe you can tell me. My goal was that White would chase Wins, Black would chase Draws and Wins.