Re: [gnugo-devel] First Monte Carlo code ready for inclusion
Ben Lambrechts wrote: This are my conclusions after running several bots on CGOS: * 1000 games per level o L1 :: 1698 o L10 :: 1813 * 8000 games per level o L1 :: 1777 o L10 :: 1938 * 10.000 games per level o L1 :: 1736 o L10 :: 1827 I don't know why the games at 10k games per level are weaker than the ones at 8k. Maybe it is because with the 10k-ones, I didn't rebuild gnugo with MAX_BOARD 9. Most likely because your computer isn't fast enough to support 1 games per level at CGOS time settings, causing the time management to regularly play at rather low levels. To see what's happening with the level settings you should use the -d 0x400 option (DEBUG_TIME), and make sure that stderr output appears somewhere you can read it. I managed to reach 1986 (mgtest1) with level 10, 6000 games per level, MAX_BOARD=19 and 1982 (mgtest2) with level 10, 7500 games per level, MAX_BOARD=9. This discrepancy I attribute to CGOS rating uncertainty. /Gunnar ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel
Re: [gnugo-devel] First Monte Carlo code ready for inclusion
Ben Lambrechts wrote: What command line options are you using, more exactly? tcl.exe cgos3.tcl GnuCvs-1-MC * gnugoMC.exe --mode gtp --cache-size 512 --chinese-rules --level 1 --capture-all-dead --monte-carlo tcl.exe cgos3.tcl GnuCvs-10-MC * gnugoMC.exe --mode gtp --cache-size 512 --chinese-rules --level 10 --capture-all-dead --monte-carlo Ok, I sort of suspected that. The only practical difference between these is that the first will start out at level 1 and play the first few moves a little faster until it reaches level 10. This may be slightly beneficial since the first few moves usually come from the fuseki database anyway and the time saved can be better used later. On the other hand it can't be very much time since move generation in the fuseki is rather fast. To reduce the span of levels that the time management is allowed to use, you need to set --max-level (default 10) and/or --min-level (default 0) options. The proper change above is to add --max-level 1 for GnuCvs-1-MC. Additionally, on CGOS you should use --positional-superko to avoid occasional losses for illegal moves. If you want debug output of the level changes done by the time management, add -d0x400. For maximum 9x9 strength you should also change MAX_BOARD in engine/board.h to 9 and rebuild, then set --mc-games-per-level as high as you can afford without the level dropping frequently due to shortage of time. /Gunnar ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel
Re: [gnugo-devel] First Monte Carlo code ready for inclusion
So the correct commands are: (after I rebuild gnugo with modified board.h) tcl.exe cgos3.tcl GnuCvs-1-MC * gnugoMC.exe --mode gtp --cache-size 512 --chinese-rules _--max-level 1_ --capture-all-dead _--positional-superko_ --monte-carlo _--mc-games-per-level TOTEST_ and tcl.exe cgos3.tcl GnuCvs-10-MC * gnugoMC.exe --mode gtp --cache-size 512 --chinese-rules --level 10 --capture-all-dead _--positional-superko_ --monte-carlo _--mc-games-per-level TOTEST_ Correct me if I am wrong please Ben tcl.exe cgos3.tcl GnuCvs-1-MC * gnugoMC.exe --mode gtp --cache-size 512 --chinese-rules --level 1 --capture-all-dead --monte-carlo tcl.exe cgos3.tcl GnuCvs-10-MC * gnugoMC.exe --mode gtp --cache-size 512 --chinese-rules --level 10 --capture-all-dead --monte-carlo Ok, I sort of suspected that. The only practical difference between these is that the first will start out at level 1 and play the first few moves a little faster until it reaches level 10. This may be slightly beneficial since the first few moves usually come from the fuseki database anyway and the time saved can be better used later. On the other hand it can't be very much time since move generation in the fuseki is rather fast. To reduce the span of levels that the time management is allowed to use, you need to set --max-level (default 10) and/or --min-level (default 0) options. The proper change above is to add --max-level 1 for GnuCvs-1-MC. Additionally, on CGOS you should use --positional-superko to avoid occasional losses for illegal moves. If you want debug output of the level changes done by the time management, add -d0x400. For maximum 9x9 strength you should also change MAX_BOARD in engine/board.h to 9 and rebuild, then set --mc-games-per-level as high as you can afford without the level dropping frequently due to shortage of time. /Gunnar ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel
Re: [gnugo-devel] First Monte Carlo code ready for inclusion
If MC is implemented around GNU Go I thing that speed is more important then the quality of the moves. That way low level is maybe better then high level. MC is very dependent on speed. Jens I've posted a new patch at http://trac.gnugo.org/gnugo/ticket/150 which contains a first version of my Monte Carlo code which should be in good enough shape to be added to CVS. It doesn't represent the full strength of the MonteGNU account on CGOS 9x9 but should at least suffice for a 2000 rating there. To turn it on, use the option --monte-carlo. To set the number of Monte Carlo simulations per move, use the option --mc-games-per-level. That number (default 8000) is multiplied by the current level to give the actual number of simulations per move. Notice, however, that the --monte-carlo option will set the maximum allowed boardsize to 9x9. This is intentional because the code simply doesn't scale to larger boards. Of course anybody who is competent enough to modify the source code can change this, but I think that's a fair barrier. Please try it out. Now in CVS. I've put the CVS version on CGOS this weekend. (GnuCvs-1-MC and GnuCvs-10-MC) The strange thing about this is that it plays better in level 1 than in level 10. Ben ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel
Re: [gnugo-devel] First Monte Carlo code ready for inclusion
If MC is implemented around GNU Go I thing that speed is more important then the quality of the moves. That way low level is maybe better then high level. MC is very dependent on speed. Jens I've posted a new patch at http://trac.gnugo.org/gnugo/ticket/150 which contains a first version of my Monte Carlo code which should be in good enough shape to be added to CVS. It doesn't represent the full strength of the MonteGNU account on CGOS 9x9 but should at least suffice for a 2000 rating there. To turn it on, use the option --monte-carlo. To set the number of Monte Carlo simulations per move, use the option --mc-games-per-level. That number (default 8000) is multiplied by the current level to give the actual number of simulations per move. Notice, however, that the --monte-carlo option will set the maximum allowed boardsize to 9x9. This is intentional because the code simply doesn't scale to larger boards. Of course anybody who is competent enough to modify the source code can change this, but I think that's a fair barrier. Please try it out. Now in CVS. I've put the CVS version on CGOS this weekend. (GnuCvs-1-MC and GnuCvs-10-MC) The strange thing about this is that it plays better in level 1 than in level 10. Ben ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel
Re: [gnugo-devel] First Monte Carlo code ready for inclusion
Sorry for my flooding the mailing list. This stupid mailer and the slow computer just killing me sometimes. Jens If MC is implemented around GNU Go I thing that speed is more important then the quality of the moves. That way low level is maybe better then high level. MC is very dependent on speed. Jens I've posted a new patch at http://trac.gnugo.org/gnugo/ticket/150 which contains a first version of my Monte Carlo code which should be in good enough shape to be added to CVS. It doesn't represent the full strength of the MonteGNU account on CGOS 9x9 but should at least suffice for a 2000 rating there. To turn it on, use the option --monte-carlo. To set the number of Monte Carlo simulations per move, use the option --mc-games-per-level. That number (default 8000) is multiplied by the current level to give the actual number of simulations per move. Notice, however, that the --monte-carlo option will set the maximum allowed boardsize to 9x9. This is intentional because the code simply doesn't scale to larger boards. Of course anybody who is competent enough to modify the source code can change this, but I think that's a fair barrier. Please try it out. Now in CVS. I've put the CVS version on CGOS this weekend. (GnuCvs-1-MC and GnuCvs-10-MC) The strange thing about this is that it plays better in level 1 than in level 10. Ben ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel
Re: [gnugo-devel] First Monte Carlo code ready for inclusion
Ben Lambrechts wrote: I've posted a new patch at http://trac.gnugo.org/gnugo/ticket/150 which contains a first version of my Monte Carlo code which should be in good enough shape to be added to CVS. It doesn't represent the full strength of the MonteGNU account on CGOS 9x9 but should at least suffice for a 2000 rating there. To turn it on, use the option --monte-carlo. To set the number of Monte Carlo simulations per move, use the option --mc-games-per-level. That number (default 8000) is multiplied by the current level to give the actual number of simulations per move. Notice, however, that the --monte-carlo option will set the maximum allowed boardsize to 9x9. This is intentional because the code simply doesn't scale to larger boards. Of course anybody who is competent enough to modify the source code can change this, but I think that's a fair barrier. Please try it out. Now in CVS. I've put the CVS version on CGOS this weekend. (GnuCvs-1-MC and GnuCvs-10-MC) The strange thing about this is that it plays better in level 1 than in level 10. What command line options are you using, more exactly? /Gunnar ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel
Re: [gnugo-devel] First Monte Carlo code ready for inclusion
What command line options are you using, more exactly? tcl.exe cgos3.tcl GnuCvs-1-MC * gnugoMC.exe --mode gtp --cache-size 512 --chinese-rules --level 1 --capture-all-dead --monte-carlo tcl.exe cgos3.tcl GnuCvs-10-MC * gnugoMC.exe --mode gtp --cache-size 512 --chinese-rules --level 10 --capture-all-dead --monte-carlo ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel
Re: [gnugo-devel] First Monte Carlo code ready for inclusion
I've posted a new patch at http://trac.gnugo.org/gnugo/ticket/150 which contains a first version of my Monte Carlo code which should be in good enough shape to be added to CVS. It doesn't represent the full strength of the MonteGNU account on CGOS 9x9 but should at least suffice for a 2000 rating there. To turn it on, use the option --monte-carlo. To set the number of Monte Carlo simulations per move, use the option --mc-games-per-level. That number (default 8000) is multiplied by the current level to give the actual number of simulations per move. Notice, however, that the --monte-carlo option will set the maximum allowed boardsize to 9x9. This is intentional because the code simply doesn't scale to larger boards. Of course anybody who is competent enough to modify the source code can change this, but I think that's a fair barrier. Please try it out. Now in CVS. I've put the CVS version on CGOS this weekend. (GnuCvs-1-MC and GnuCvs-10-MC) The strange thing about this is that it plays better in level 1 than in level 10. Ben ___ gnugo-devel mailing list gnugo-devel@gnu.org http://lists.gnu.org/mailman/listinfo/gnugo-devel