|
From: | Philippe Michel |
Subject: | Re: [Bug-gnubg] Is GNUBG actively developed? |
Date: | Thu, 7 Feb 2013 00:04:42 +0100 (CET) |
User-agent: | Alpine 2.00 (BSF 1167 2008-08-23) |
I am willing to believe the new nets are better, but I have not seen the results of a long-enough/statistically-significant run of matches between the old and new.
I have uploaded my latest nets at http://www.gnubg.org/media/nn-training/pmichel/nets/20130121/
Access doesn't seem to work well for now, apparently as a consequence of the site hacking. Direct links to the files should be fine :
http://www.gnubg.org/media/nn-training/pmichel/nets/20130121/gnubg.wd.bz2 http://www.gnubg.org/media/nn-training/pmichel/nets/20130121/gnubg.weights.bz2 http://www.gnubg.org/media/nn-training/pmichel/nets/20130121/nngnubg.weights.bz2 Matches against gnubg 0.90 nets, using matchplay.py -m 5 --q1 ... give : - at 0 ply : # X - 502560 O - 497440 (50.26-%) (268114 117223 114663) # X - 501845 O - 498155 (50.18+%) (267521 117162 115317) # X - 501867 O - 498133 (50.19-%) (268461 116703 114836) # X - 502754 O - 497246 (50.28-%) (268354 117200 114446) tallying to X 2009026 O 1990974 (50.226%) and a sd of ca. 0.048% - at 2 ply : # X - 1253 O - 1247 (50.12+%) (723 265 262) # X - 1305 O - 1195 (52.20+%) (719 293 238) # X - 1313 O - 1187 (52.52+%) (675 319 256) # X - 1290 O - 1210 (51.60+%) (710 290 250) # X - 1266 O - 1234 (50.64+%) (706 280 264) # X - 1290 O - 1210 (51.60+%) (710 290 250) # X - 1299 O - 1201 (51.96+%) (703 298 249) # X - 1258 O - 1242 (50.32+%) (670 294 286) # X - 1306 O - 1194 (52.24+%) (716 295 239) # X - 1275 O - 1225 (51.00+%) (641 317 292) # X - 1249 O - 1251 (49.96+%) (679 285 286) # X - 1258 O - 1242 (50.32+%) (700 279 271) # X - 1253 O - 1247 (50.12+%) (683 285 282) # X - 1249 O - 1251 (49.96+%) (733 258 259) # X - 1262 O - 1238 (50.48+%) (674 294 282) # X - 1271 O - 1229 (50.84+%) (701 285 264) # X - 1259 O - 1241 (50.36+%) (701 279 270) # X - 1263 O - 1237 (50.52+%) (697 283 270) # X - 1254 O - 1246 (50.16+%) (678 288 284) # X - 1292 O - 1208 (51.68+%) (684 304 262) for X 25465 O 24535 (50.930%) and a sd of 0.42%The former is comfortably significant, the latter looks, at worst, encouraging. FWIW, the sets of 2500 were taking about a day on my machine and I guess someone dedicating a Core i5 CPU to this should be able to get 100000 matches per week.
[Prev in Thread] | Current Thread | [Next in Thread] |