Computer eval of Fischer-Random (960) starting positions?

@dboing said in #29:

I wish we could have more rational discussions with more information about what is in the setups and how that affects the thinking or learning problem.. Not just envelope statistics.

Many people, (myself included) say that puzzles are puzzles and games are games? I think this sort of applies to Fischer Random. We all try to look at starting positions as if there's some profound plan we see by looking at them. Then we see that puzzles are not games, and there is way more to it than just looking at an eval/fen for that starting position. Even the results shown in those tables linked early in the thread show the starting eval doesn't translate well to game results. The best eval isn't even close to showing the highest win rate for white. In some ways this doesn't surprise me. For the vast majority of people, 960 is like communicating in a language we don't speak.

If one thing is a constant across all positions, Magnus seems to dominate regardless of the back rank. He's pretty good at 518, and looks even stronger when the position !=518. What this tells me is, something about FR/960 does not follow conventional chess logic. Or if it does, we practice the wrong logic. He has insights no other player possesses. You don't go 9/9 like he just did against the worlds top players, without some unique perspective. I don't know where to begin the learning process.

For now, I'm happy with envelope statistics.

@collective_enjoyer said in #25:

I value counterattacking chances as black and always create asymmetry when I can, but I am genuinely confused how black can have a perceived advantage at the starting position.

Yes, I also bugged a bit on this one.. But in distribution over all back rank setups, having differring engine scoring and given a margin of engine is not truth error (blasphemy?) , that it might vary already, might make it so that it could actually reduce the first turn advance.... That! I would find possible, since the question is of color axis imbalance (what else could it be? Ask Flores book), that it might end up nibbling on the white initial advantage of the most symetric of all 960s, the standard one (it has other properties, but in some way I am not sure it is only about color axis "symetry", or lateral symetry of the minor pieces or rook, the RNB permutation iself might be having properties: did this 3 piece (or mobility rulesubset) ever change over chess (not just modern, but all the way back to Indian first species of chess)

Did I not see some small in the decimal significant digits, negative.. Is this within error range of the mighty engine oracles?

In some sense it makes sense that it would not be a completely invertible SF score on the negative side.. (distribution of engine score over the domain of back ranks, since that is not ordered we might just consider histogram on the SF score axis.

but I still wonder of the specific board arguments that would negate the first side advantage.... how could it be..

that would be chess theory worthy to discuss, no? not engine stuff at first (it might end up in engine point of view entrails eventually), but is there any chess logic argument based discussion to be had about how it could happen?

Yes, good point, thanks for confirming this, currently dissonant bit of information (currently dissonant, up for discussion).

Also sorry if some of my questions have already been answered.. I tend to skip things haphazardly, like starting from the end.. or reading fast... perhaps we could make visual graphical summaries of what has transpired? so far.

then some of us might want to trace back into the 4 pages where this was discussed. I know, seems lazy on my part.

@dboing You're fine boss I actually really liked thinking about this.

For those curious I used Stockfish 17 on my own computer to run at a depth of 40 for 25 seconds for every starting position on my own personal computer. Advantage perceived here could be a limitation of compute time and depth, but my intuition wants to take a closer look. The SPs 30, 110, 433, 744 were what I was able to have a hit for some perceived advantage for black. What these 4 positions have in common are knights that start on the same color complexes that are a square apart, and a queen that is not on either d or e files. 110, 433, and 744 also have initially unprotected e or d pawns.

Interestingly, SP 404 recommended by @AsDaGo doesn't share those features, but black is doing well statistically and I'm pondering on what other things are tricky for white to stay up in tempo and solidify a better position.

As of 05.28.2025, from these 5 positions SP 110 scores the best on the lichess database White 49% 3% draw 48% Black after 23,815 games.

As for the game sample sizes what would be meaningful to include? Should we be discarding games of a certain Elo or are those games also informative?