💡📝H. G. Muller wrote on Tue, Jun 8, 2021 11:16 AM EDT:
I made some fixes to the diagram. The SAN parser (used when you paste a game into the diagram) choked on moves with a Shogi-style piece id +L, because it tried to handle Shogi promotion suffixes and matching the piece id in the wrong order. It coulld also not understand Chess-style promotion sufixes with Shogi piece names ("=+L"), and unfortunately this is how the SAN generator wrote Shogi promotions. I now altered the latter to just append a + when the promoOffset diagram parameter is defined as non-zero, and the difference beween the promotion piece and the oriinal piece is equal to it. And the parser now understands the + suffix as promotion when promoOffset is non-zero. (Otherwise it inores it, assuming it means 'check'.)
I also made a small improvement to the AI; this searches novel moves deeper than moves that already were possible for the same player in his previous turn. But it has to know the previous two half-moves to determine whether moves are new, and in the root this was not the case. I now let it remember the previous move it played, for that purpose. As a consequence it will now consumate mate-in-1 threats it created with its previous move even on the lowest (2-ply) setting.
I also added a kludge to combat repetition: moves with the same piece as before now are discouraged by a light penalty, and when the move goes back to the square it came from (and nothing was captured with these two moves), it will be penalized with 1/4 of its advantage (if it has any).
I made some fixes to the diagram. The SAN parser (used when you paste a game into the diagram) choked on moves with a Shogi-style piece id +L, because it tried to handle Shogi promotion suffixes and matching the piece id in the wrong order. It coulld also not understand Chess-style promotion sufixes with Shogi piece names ("=+L"), and unfortunately this is how the SAN generator wrote Shogi promotions. I now altered the latter to just append a + when the promoOffset diagram parameter is defined as non-zero, and the difference beween the promotion piece and the oriinal piece is equal to it. And the parser now understands the + suffix as promotion when promoOffset is non-zero. (Otherwise it inores it, assuming it means 'check'.)
I also made a small improvement to the AI; this searches novel moves deeper than moves that already were possible for the same player in his previous turn. But it has to know the previous two half-moves to determine whether moves are new, and in the root this was not the case. I now let it remember the previous move it played, for that purpose. As a consequence it will now consumate mate-in-1 threats it created with its previous move even on the lowest (2-ply) setting.
I also added a kludge to combat repetition: moves with the same piece as before now are discouraged by a light penalty, and when the move goes back to the square it came from (and nothing was captured with these two moves), it will be penalized with 1/4 of its advantage (if it has any).