I looked for obvious reasons at the position as well and I did a few games by hand and in 6 of 6 games I closed out both checkers (assisted by BGBlitz).
because I used the forthcoming TachiAI6 and not the current production TachiAI5 I checked the evaluation of TAI5 and on lower levels it sees a bit higher winning probabilities for the 15 checker player (good!) and the rollout was 55.54% so nearly the same value.
Having XG an 51% (Stick in the post here) instead of the 35% (3-ply move; XG+ cube; normal filters) is clearly attributed to the gigantic filter. More on this later.
When there is a 20% difference there must be a reason (and the 2-checker player can’t be misplayed). So I thought let XG play some games against itself with XG++ from this position and have a look
But it turned out that I only played 1 game because even XG++ was IMHO far off (Here is the match: https://www.bgblitz.com/download/picsForOthers/FB_Containment_1.mat)
I looked at the first 4 moves:
XGID=----B-Ca----ABAa----A-C-AA:1:1:1:65:0:0:3:0:10
XG chooses Bar/20, 12/6
BGBlitz chooses
b/19, 20/15* even with 1-ply
BGBlitz regards Bar/20, 12/6 as a 0.347 error
If you force xg to XG++ b/19, 20/15* it comes out as best.
XGID=----B-D--a---BA----aB-C-A-:1:1:1:54:0:0:3:0:10
XG chooses 24/19*, 14/10;
BGBlitz chooses
24/19*, 13/9* and 24/19*, 22/18 being close;
BGBlitz regards 24/19*, 14/10 as an 0,074 error
XGID=a---B-D--aA--B-----AB-C---:1:1:1:54:0:0:3:0:10
XG chooses 19/14, 10/6
BGBlitz chooses
22/17,13/9* or 20/15, 13/9*
BGBlitz regards 19/14, 10/6 as an 0.227 error
If you force xg to XG++ 22/17,13/9* it comes out as best and 19/14,10/6 is an 0.173 error
XGID=a---B-E--a---BA-----B-C---:1:1:1:41:0:0:3:0:10
XG chooses 22/21, 22/18
BGBlitz chooses 22/21, 13/9* from 2-ply and the to me obvious 13/9*,6/5 is close from 5-ply on
22/21, 22/18 is only an 0.039 error
If you force xg to XG++ 22/21, 13/9* and 13/9*,6/5 it comes out as 1st and 3rd.
I stopped at that point. What is obvious to me is that on 1-ply XGs understanding is too bad so that reasonable moves are kicked out and not analyzed in higher plies (best move regarded as up to an 0.600 error at 1-ply). Being at top when forced to be analyzed shows that. But that helps not always (2nd move). Again: this game was played with XG++!
Given that the moves of BGBlitz are pretty reasonable (if 13/9,6/5 in the 4th move would have come earlier than 5-ply I would say *very* reasonable) it seems another example to support my claim that BGBlitz has an edge over XG in containment games :)
Feedback appreciated.