TalkChess.com

[Gerd Isenberg]

Hi Rémi,

Thanks for providing the paper, but sorry for my lack of understanding - I am a novice in statistics as well in neuronal nets, temporal differences and such stuff. I read your paper a few times - and sometimes there was a kind of flighty mental dust which gives me the illusion to understand something

A few questions related to the domain specific features. You somehow index the database of your learned values by a move - like indexing a history table? Each feature has a kind of let say persistant history table with former learned values. And if featuers of a move have in combination (weighted by game-state?) a lower probability than some threshold, it is a candidate for reduction or even pruning. Does that roughly describe how it works?

What features would you suggest for the domain of chess for your algorithm? Are there disjoint features considering game state, closed/open positions, blocked pawn center, opposite castles, etc.? According to your distance features - how would those translate to chess since in Go you have only to-squares, but not from-to. Is it something like center- or piece-king tropism?

Thanks for your patience,
Gerd