On Sun, Jun 19, 2011 at 4:48 PM, Michael Williams < [email protected]> wrote:
> I get it. But ssh access is a barrier to participation. Even though I want to do this study, I don't have the time or energy to build a distributed auto-tester just for this "one-off" experiment. Windows is horrible for one-off experiments like this but Linux is built around useful abstractions and tools that can be combined in useful ways. I'm just trying to find a way that makes it possible to organize this experiment without requiring me to donate 200 man hours of my time. I'm actually thinking about the CGOS client now. I talked about how CGOS itself is not suited to this experiment, but I might be able to modify the simple CGOS client to connect to a local process that pretends it's a GTP Go program and so it would run on any OS. I'm also open to suggestions. Don > On Sat, Jun 18, 2011 at 7:58 PM, Don Dailey <[email protected]> wrote: > >> Probably the easiest thing for me is to run the programs via remote shell >> (ssh) using my own self-tester. I don't remember how my self-tester works >> but my chess and shogi self-tester multiplexes games, so it's a trivial >> setup, except that I would have to make the modification to restrict the >> number of connections to any particular machine (and this would need to be >> configurable.) >> >> The reason for this is that my tester(s) assume everything is running on >> one machine (and I specify how many simultaneous games to allow) but if a >> given player has not played as many games as another player, it would >> schedule several games with that one player. But it would not be good if >> 20 instances of "[email protected] /home/joe/bin/zen -level 10000" needed >> to be scheduled since they would all be running on the same machine. If >> it were not for that, I would not need to make ANY modifications to the >> tester, I could just configure each player using that form assuming I set >> up the ssh keys properly and have an account on foobar.com. >> >> Another problem with the autotester is that it only knows how to play >> round robin matches, but that is not a big problem since I can partition >> the test easily enough. What I would do is run a full round robin just >> long enough to get reasonably stable ratings, then based on the ratings I >> could partition the players into separate sections by strength. And from >> time to time I would re-calibrate the ratings and change the partitions >> sizes in order that all the players have good interaction are not playing in >> isolated rating pools. >> >> I have not looked too closely but if my go tester does not multiplex games >> I can still simply run several instances of it. I just want to find a >> setup that does not require too much babysitting on my part. >> >> Don >> >> >> >> >> >> On Sat, Jun 18, 2011 at 7:08 PM, Michael Williams < >> [email protected]> wrote: >> >>> You could write a new server that uses a CGOS-compatible interface so >>> that the CGOS clients can be used. The server would always match-up two >>> engines from the same IP address that start with the same name prefix. And >>> would only run one game at a time for that IP address and prefix. On the >>> client side, you would run several instances for each CPU core, but only two >>> of them would be in a game at the same time (against each other). >>> >>> Or something like that. >>> >>> >>> >>> On Sat, Jun 18, 2011 at 6:50 PM, Don Dailey <[email protected]>wrote: >>> >>>> >>>> >>>> On Sat, Jun 18, 2011 at 6:25 PM, Jacques Basaldúa <[email protected]>wrote: >>>> >>>>> 1. Perhaps Erica may be of help. 10 months ago it >>>>> won the Olympiad, so it should be strong enough. If >>>>> Aja is so kind to let us copy of the binary. >>>>> >>>>> 2. I can donate cpu, but it will be better in 3 weeks >>>>> when I return from a short holiday and have a new >>>>> machine more. >>>>> >>>>> 3. Maybe the idea of using a special CGOS makes it >>>>> easier for everybody. >>>>> >>>> >>>> I thought of CGOS, but CGOS is a terrible solutions for this specific >>>> thing because it would have to be configured for a really long time control >>>> in order to avoid time forfeits. But it will not schedule a round until >>>> all games are complete from the previous round so most computers would be >>>> idle most of the time. It would be better to have a system that keeps all >>>> computers busy all of the time. >>>> >>>> On the other hand, that is a pretty simple way to do it but don't know >>>> if we would have the patience for it ... >>>> >>>> The previous study was good because it represented a huge amount of CPU >>>> effort, it was not just running a few games for a couple of days but it >>>> was >>>> thousands of games played on 40 or 50 cores over a period of weeks. I >>>> don't think CGOS would be good for this. >>>> >>>> I'm still thinking about how it might be done without a huge amount of >>>> effort on my part - I would really like to do this study. >>>> >>>> Let me ask the group this question: How do you run automated testing >>>> under CGOS conditions (other than using CGOS?) What tools are available >>>> that work under Linux? >>>> >>>> Don >>>> >>>> >>>> >>>> >>>> >>>>> >>>>> >>>>> Jacques. >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> >>>>> ______________________________**_________________ >>>>> Computer-go mailing list >>>>> [email protected] >>>>> http://dvandva.org/cgi-bin/**mailman/listinfo/computer-go<http://dvandva.org/cgi-bin/mailman/listinfo/computer-go> >>>>> >>>> >>>> >>>> _______________________________________________ >>>> Computer-go mailing list >>>> [email protected] >>>> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >>>> >>> >>> >>> _______________________________________________ >>> Computer-go mailing list >>> [email protected] >>> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >>> >> >> >> _______________________________________________ >> Computer-go mailing list >> [email protected] >> http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >> > > > _______________________________________________ > Computer-go mailing list > [email protected] > http://dvandva.org/cgi-bin/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
