Opening Novelty by Neural Networks

Sort:
drmrboss

Legendary Stockfish running on 90 cores CPU got outplayed on Scandi opening.

Details https://www.chess.com/forum/view/chess-equipment/advanced-attacking-books?page=3

drmrboss
NMinSixMonths wrote:

What is 81.Qd4+ about? 81.Qh4+ is mate in two.

It is common pattern in neural networks of current training system where they give away pieces in endgame!!

 Why?

NN are trained with winning percentage.  If there are two moves between  99%winning vs 99.99% winning , they will choose 99.99%.

Having more pieces will give you the higher risk of stalemate and reduce the chance of winning.

For example, if you have KQR vs K, there is more risk of stalemate than KR vs K, so NN  will give away Queen to reduce risk of draws( stalemates) and increase the chance of winning.

 

During training there is no reward for winning mate in 2 vs mate in 20. ( NN doesnt know anything about chess except basic rules and in chess rule there is no additional reward for shorter win). They become expert by choosing the moves that give them better chance of winning during training.