Some experimental "similarity" data
Re: Some experimental "similarity" data
For some methods/problems with making hierarchal classifications from this sort of data, I might refer to http://www.let.rug.nl/~nerbonne/teach/r ... c-2009.pdf particularly starting around page 22/23.
Re: Some experimental "similarity" data
It remains to be seen if this is really a viable possibility, but it might be that some of the ideas here could be used to detect "easy" moves. The premise here is that if enough tweaks of your algorithm (such as removing null move) all give the same "bestmove" after a short search, this is good evidence that the move is "easy". Indeed, it seems to me that one thing that the experiment shows is that a large percentage of moves (over 50%) are "easy" in some sense, as dramatically changing the search, or doubling or quadrupling the time, doesn't lead to great variance in "bestmove" results.
So here would be a simplistic idea. For the first ~10% of the alloted time, run (say) 8 search variants in parallel, and if 7+ of them give the same "bestmove" (and perhaps some other conditions are met) conclude this move is "easy" and can be played immediately. I'm not sure how this would dovetail with pondering.
So here would be a simplistic idea. For the first ~10% of the alloted time, run (say) 8 search variants in parallel, and if 7+ of them give the same "bestmove" (and perhaps some other conditions are met) conclude this move is "easy" and can be played immediately. I'm not sure how this would dovetail with pondering.