recently i let Sapling v1.1.2 play a 5m3s game with Patricia, the new 'super agressive' engine, also with NN .. at one moment in the game Sapling was clearly better (+3) but for almost 50 moves no piece was captured nor pawn moved, so this would lead to a draw and Sapling decided to push a pawn, avoid the 50-move rule because it was winning and it 'thought' the remaining position had a lower eval (+2.5) but enough not to lose, or so .. against Patricia this didn't work out .. she provoked the complications, which were indeed hard to cope ..
at the end, in theory White was much better, but how ?
i have no clue how my description of this 'behaviour' can help you improve the engine, but here's the (long) game : https://lichess.org/4uUXF7xC
note: the Stockfish eval values on LiChess are much lower then Sapling' judgements
FYI - Sorry it's taken me a while to respond! Been very busy the last few days, will take some time tomorrow to get around to replying to all your feedback / insights!
recently i let Sapling v1.1.2 play a 5m3s game with Patricia, the new 'super agressive' engine, also with NN .. at one moment in the game Sapling was clearly better (+3) but for almost 50 moves no piece was captured nor pawn moved, so this would lead to a draw and Sapling decided to push a pawn, avoid the 50-move rule because it was winning and it 'thought' the remaining position had a lower eval (+2.5) but enough not to lose, or so .. against Patricia this didn't work out .. she provoked the complications, which were indeed hard to cope ..
at the end, in theory White was much better, but how ? i have no clue how my description of this 'behaviour' can help you improve the engine, but here's the (long) game : https://lichess.org/4uUXF7xC
note: the Stockfish eval values on LiChess are much lower then Sapling' judgements