A good thing I didn't get overexcited, the result of my test is too close to call:Don wrote:Yes, this can be embarrassing. I sometimes call Larry and say, "it's off to a great start!!!" only be made a fool out of. Of course I know better but you want to see improvement so you get excited.lucasart wrote: Now for using the refutation in LMR, I will simply try to use it as an LMR veto (like killers). Given the way DiscoCheck does search reductions (not quite standard and probably lame) reducing the reduction by one ply as you suggest, or not reducing at all the refutation move, would amount to the same.
But when using the refutation move as an LMR veto, I will use (2), because the fact that it's more specific probably means the cost is going to be less. For the moment it's looking very good, but I've just started the test, so I'm a bit worried to post early results and get too excited
* 10k games in 10"+0.1"
* W: 2819 L: 2760 D: 4421, score = 50.3%, p-value = 78.5%
Anyway, I'll still commit, as it is slightly positive (even if within error bar) and more importaltly it reduces reductions, hence the risk of tactical accidents (which seem to be the cause of lots of losses at long TC when looking at Graham's tournaments).
This counter move heuristic is a real gem. Many things I need to try in DiscoCheck:
1. Joona Kiiski's idea:
1.0 using two slots instead of 1
1.1 why not N slots and seeing which N is best?
1.2 with that can the difference fsq/tsq instead of piece/tsq be finally measurable?
2. Daniel Homan's idea
2.0 using a double move refutation hash table (nice generalisation of the counter move heuristic)
2.1 why not N consecutive moves? (just as easy to implement with the added detail that if N is odd you must xor out the zobrist turn of play). Intuitively it seems unlikely that extending it t more than 2 moves will work, but te only way to know for sure is to test.
2.1 same question as for the basic counter move heuristic: use in LMR or just in ordering? (my guess is that doing both will be best in DiscoCheck)