Author: Michael Henderson
Date: 09:16:32 09/12/04
Go up one level in this thread
On September 12, 2004 at 07:43:24, Kim Roper Jensen wrote: >On September 11, 2004 at 19:43:56, Stuart Cracraft wrote: > >>Hi, >> >>Added fractional extensions today and it boosted solution rate >>a little more than 1% for the problem set. This was just a set >>I threw in without much thought so that's pretty good considering >>improvements lately have been hard to come by though the board >>is helping to drive them a lot faster than if I were working it >>alone. >> >>I initialize variable "extend" to 0.75 on entry to main search but I'll >>probably make that 0 and the next number below a 1. I don't keep >>track of the number of checks in a row but could to do the skip >>so this isn't like Bob's. The 0.75 is just homage. >>If in check, add 0.75. >>If a recapture, add 0.75. >>If a pawn extension, add 0.50 (this one is disabled presently but now >> that fractional extensions are in, it and the other extensions I haven't >> left enabled for fear of blowup can be all added in most likely.) >>If extension total is >= 1.0, depth+=1 otherwise keep previous setting of depth. >>... various non-extensions work here... test whether depth>=0 for >> call to quiescence, null move with adaptive null move, etc. >> (verified null move doesn't work as well for me for my short searches) >>Generate moves, score, and sort >>If only one legal move, give a free extension beyond the above. >> >>So, an example. If in check, extend by 1 (since 1.50 (0.75+0.75)). >>If in check and this node is after a recapture extend by 1 (0.75+0.75+0.75>=1) >>If in check and only one legal move out extend null move by 1 but then >> 1-N moves by 2 (0.75+0.75>=1 + one legal move (1)) >> >>I was surprised to get a jump like this on the first attempt addition >>of the fractional extension feature. 99/100 times a new thing added drops >>solution rate slightly or more rarely by a lot. >> >>with fractional extensions >>+ 6.19/24.31 82% 246/300 249.13 58736604 195789/1/235770 0/639780/1159759/415210 >>/12993546/0 >>pawnx/recapx/qcheckx/checkx/futilx/onereplyx >>Missed: >>2 21 46 71 80 86 87 92 99 100 116 128 131 141 145 155 163 178 180 196 200 210 21 >>3 214 221 222 223 226 228 229 230 237 241 243 245 247 252 255 256 261 262 265 26 >>6 269 274 282 283 287 288 291 293 296 297 299 >>branching factor = 4.06 >> >>before fractional extensions >>+ 6.37/23.52 80% 242/300 248.87 58082496 193608/1/233381 0/640257/1159613/384586 >>/13051686/0 >>pawnx/recapx/qcheckx/checkx/futilx/onereplyx >>Missed: >>2 21 49 55 71 80 86 87 92 97 99 100 116 128 131 139 141 145 155 163 180 196 200 >>207 210 213 221 222 223 226 228 229 230 231 237 241 243 245 247 250 252 255 256 >>260 261 262 265 269 274 282 283 287 288 291 293 296 297 299 >>branching factor = 3.81 >> >>qcheckx refers to evading checks in the quiescence. This is just a measure >>of the handoffs to the main search from a quiescence that is entered in >>check, using depth=1 -- in these cases, the check extension in the main >>search is disabled and not used since it already received an extra ply >>from the handoff. >> >>Curious what are all the different fractional extensions people are using >>for the various types of extensions. >> >>Now's your chance to write all of yours in here. E-X-T-E-N-D yourself! >> >>Stuart > >Hi Stuart > >Being a not so good programmer myself with a very nonfunctional program I just >remembered someone mentioned sometime ago doing negative extensions also for >reducing tree size > >/kim the classic name for this stategy is naturally called, "reductions" :) i have never tried them myself unless null move counts
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.