Re: Re:[computer-go] Time Control for the new CGOS
On Tue, Mar 27, 2007 at 08:43:23PM -0400, Don Dailey wrote: On Tue, 2007-03-27 at 16:02 -0700, Christoph Birk wrote: On Wed, 28 Mar 2007, Heikki Levanto wrote: P.S. How about starting a new round when (say) 75% of the players are free? That introduces a bias towards pairing faster programs against each other. I've really struggled with this one. In the end, scheduling is far easier and has far less side effects if I make them discreet. [...] A little analysis shows that this does not decrease the playing rate much at all for programs that use most of their time. For the really fast programs, you will clearly get scheduled less often. I usually make decisions, where there is an issue, in favor of the stronger programs as long as it doesn't introduce gross unfairness to the weaker programs.In this case I don't want to introduce a scheduling algorithm that encourages random players to play zillions of games. Fair enough, it was just a lose thought - Heikki -- Heikki Levanto In Murphy We Turst heikki (at) lsd (dot) dk ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] Time Control for the new CGOS
From: Don Dailey [EMAIL PROTECTED] I am considering to change the time control when I change over officially to 5 minutes instead of 10. 5 minutes seems more than adequate for the Monte Carlo programs which play quite strongly even at 2 minutes per game. Hello all, I prefer very short time controls, like 1 or 2 minutes for 9x9 and 5 to 10 minutes for 19x19, to obtain faster testing speeds. Isn't it true that most of us will test our (MC based) programs against opponents with fast time controls to get an accurate rating fast enough? And if so, why not extending this fast rates to CGOS? Just my thoughts, maybe I am wrong about what others do. What do the other readers think of this? Edward. _ Geen zin in typen? Praten gaat makkelijker. Nu GRATIS bellen! http://get.live.com/messenger/overview ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
RE: [computer-go] Time Control for the new CGOS
On Wed, 2007-03-28 at 12:25 +0200, Edward de Grijs wrote: From: Don Dailey [EMAIL PROTECTED] I am considering to change the time control when I change over officially to 5 minutes instead of 10. 5 minutes seems more than adequate for the Monte Carlo programs which play quite strongly even at 2 minutes per game. Hello all, I prefer very short time controls, like 1 or 2 minutes for 9x9 and 5 to 10 minutes for 19x19, to obtain faster testing speeds. Isn't it true that most of us will test our (MC based) programs against opponents with fast time controls to get an accurate rating fast enough? And if so, why not extending this fast rates to CGOS? Just my thoughts, maybe I am wrong about what others do. What do the other readers think of this? Edward. One has to strike a balance. CGOS isn't just for pure testing, it can be used to test algorithms that may require time to be realized. Or it can be viewed as a competition like a ladder. Also, the server will be visual soon, allowing users to watch games while they are being played. In this case if the games get too fast they are difficult to digest, at least for weaker players like myself. The quality is also affected by the time control and I think it's desirable to raise the quality of the games. There are many programs which could maintain a 1 minute time control, and these programs would particularly benefit (from a competetive point of view) from faster time controls and scalable algorithms would suffer accordingly. So there are pro's and con's to any time control we might consider and we should strike a reasonable balance. In some sense any time control is arbitrary - but we end up choosing time controls that are comfortable from the human point of view - in other words a human would feel relatively comfortable playing the time control - even though it's our programs playing! But I think that's good since most people develop their programs with the potential goal of playing against people. - Don _ Geen zin in typen? Praten gaat makkelijker. Nu GRATIS bellen! http://get.live.com/messenger/overview ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Time Control for the new CGOS
Hi, regarding time controls and the impossibility to please everyone, I'd like to make a suggestion: Let the engines specify a preferred time control and use a scheduler that takes that into account (as a strong recommendation). For example if there are two engines wanting to play at 10 minutes, pair them with higher probability. There could be just a few predefined time limits, maybe 5, 10, 15 minutes for 9x9. IMHO this would be better than using a different server for each time limit or than changing the time limit at intervals (although having a one week rotation between 5 and 10 minutes seems ok). Of course, the scheduling algorithm becomes more complex and it may not be as fair. On the other hand, it may not be very important to achieve fairness on a short scale as long as it is fair enough over a longer period of time. just my 2c best regards, Vlad ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Time Control for the new CGOS
Is 10 minutes a standard and if so it is standard for 19x19 or 9x9? For 19x19 I find it a little too fast. I would prefer fastest: 4 sec/move (x240 moves) = 16 min slowest: 30 sec/move (x240 moves) = 2 hours I would like to try both. Usually fast because, as you pointed, you get useful results earlier, but maybe one week each month, slow. 240 moves is above the average length of a 19x19 game. In my 50K games database it is 212 and games ending before move 100 were removed, but among weak programs it is low. Darren Cook wrote: That summarizes my main argument against short time controls: it limits the choices for experimentation so to meet time controls everyone will end up running very similar MC programs. With more time people have some breathing space to experiment with new ideas and intelligence. I agree even if MC can make very good usage of extra time, it is not the only approach that does. As said, doing *both* is also an option. Jacques. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Time Control for the new CGOS
On Tue, 2007-03-27 at 13:26 -0700, Ken Friedenbach wrote: 1 second is probably too much when network lag is not an issue, and not enough when it is... Is there some way to use ping every N moves, to get a better setting on a per connection basis? Probably, but I don't want to trust clients not to cheat by delaying their response to ping.You probably think this is incredibly paranoid of me, but if you've lived in the computer chess world you would understand why I feel this way :-) Of course you can do a statistical analysis of the response times to a number of pings and gather statistical evidence that a client is cheating or not, but this can probably be defeated with clever algorithms to vary the responses! Now I'm really being paranoid!At any rate, I don't want to get into this.I think I will just use a fixed 1/2 second forgiveness factor. - Don Ken Friedenbach On Mar 27, 2007, at 9:33 AM, Don Dailey wrote: The 2 minute server is interesting, the short time control has still allowed for very strong programs including Mogo. I am considering to change the time control when I change over officially to 5 minutes instead of 10. 5 minutes seems more than adequate for the Monte Carlo programs which play quite strongly even at 2 minutes per game. What does everything think about that? The server adds 1 second gift to each move silently, in order to allow for server overheads and network lags and glitches. This is pretty generous and actually adds a minute or more the the length of the games. I think this should probably be cut to 1/2 second instead of a full second and I will make that change. - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re:[computer-go] Time Control for the new CGOS
But what are real conditions? Is 10 minutes a standard and if so it is standard for 19x19 or 9x9? I meant for 9x9 and games against humans for example. At any rate, I will probably go with 5 minutes unless I get a lot of protests, in which case I will stay with 10 minutes. I have considered the pro's and con's below. Ok thank you. You have a stronger opinion than mine :) Sylvain ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re:[computer-go] Time Control for the new CGOS
On Tue, 2007-03-27 at 23:52 +0200, Sylvain Gelly wrote: But what are real conditions? Is 10 minutes a standard and if so it is standard for 19x19 or 9x9? I meant for 9x9 and games against humans for example. At any rate, I will probably go with 5 minutes unless I get a lot of protests, in which case I will stay with 10 minutes. I have considered the pro's and con's below. Ok thank you. You have a stronger opinion than mine :) Sylvain Actually, I only slightly prefer 5 minutes - it seems like it would be a benefit all things considered. But as I said, I'm willing to concede - I will do what the group as a whole prefers. So far nobody has spoken out in favor of 5 minutes except me - so it might end up being 10 minutes after all! - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: [computer-go] Time Control for the new CGOS
On Tue, Mar 27, 2007 at 12:33:36PM -0400, Don Dailey wrote: I am considering to change the time control when I change over officially to 5 minutes instead of 10. 5 minutes seems more than adequate for the Monte Carlo programs which play quite strongly even at 2 minutes per game. What does everything think about that? Fine by me. The server adds 1 second gift to each move silently, in order to allow for server overheads and network lags and glitches. This is pretty generous and actually adds a minute or more the the length of the games. I think this should probably be cut to 1/2 second instead of a full second and I will make that change. Quite fine by me. My quick program halgo-1.7-10k always seemed to have 120 seconds left, no matter what happened. Maybe even 0.1 seconds would be sufficient. In fact I don't quite see the point in adding the extra time. Both players get as much, and both are supposed to have the same amount of time available, so it should not make much of a difference. I can se an argument for it, in the case the opponent plays silly moves just to kill time, and hope to win on time. But we already play the games to the bitter end, so that should not matter so much. Maybe the time should only be added after a player has passed once? I am still not sure about all the consequences of this, but I don't see it as a major problem, one way or the other. - Heikki -- Heikki Levanto In Murphy We Turst heikki (at) lsd (dot) dk ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re:[computer-go] Time Control for the new CGOS
On Tue, Mar 27, 2007 at 06:15:49PM -0400, Don Dailey wrote: Actually, I only slightly prefer 5 minutes - it seems like it would be a benefit all things considered. But as I said, I'm willing to concede - I will do what the group as a whole prefers. So far nobody has spoken out in favor of 5 minutes except me - so it might end up being 10 minutes after all! Then let me cast a vote for 5 minutes - although it is a very weak vote, as I do not have a strong opinion. But the 2min tests have shown that I seem to have a powerful enough computer to do what I wanted to do... Being able to play more games in a given time is clearly a bonus. - Heikki P.S. How about starting a new round when (say) 75% of the players are free? That way, the last slow ones could skip a round, and most of the rounds would still be with most of the players. -- Heikki Levanto In Murphy We Turst heikki (at) lsd (dot) dk ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re:[computer-go] Time Control for the new CGOS
On Wed, 28 Mar 2007, Heikki Levanto wrote: P.S. How about starting a new round when (say) 75% of the players are free? That way, the last slow ones could skip a round, and most of the rounds would still be with most of the players. That introduces a bias towards pairing faster programs against each other. Christoph ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re:[computer-go] Time Control for the new CGOS
I vote for 5 minutes per side in the name of faster ratings, faster testing and faster games for casual observers to observe. Plus, computers are getting faster every day and 9x9 Go algorithms are getting better every day so it seems reasonable to speed-up the time controls from what it was when CGOS started. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re:[computer-go] Time Control for the new CGOS
On Tue, 2007-03-27 at 16:02 -0700, Christoph Birk wrote: On Wed, 28 Mar 2007, Heikki Levanto wrote: P.S. How about starting a new round when (say) 75% of the players are free? That way, the last slow ones could skip a round, and most of the rounds would still be with most of the players. That introduces a bias towards pairing faster programs against each other. Christoph I've really struggled with this one. In the end, scheduling is far easier and has far less side effects if I make them discreet. There is also a feature in the server to report informational messages - one of them reports the time until the next round, which hopefully makes the wait more bearable. Of course it's just an estimate and it's almost always overstated. A little analysis shows that this does not decrease the playing rate much at all for programs that use most of their time. For the really fast programs, you will clearly get scheduled less often. I usually make decisions, where there is an issue, in favor of the stronger programs as long as it doesn't introduce gross unfairness to the weaker programs.In this case I don't want to introduce a scheduling algorithm that encourages random players to play zillions of games. - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/ ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re:[computer-go] Time Control for the new CGOS
Either 5 or 10 minutes per side is fine by me, with a mild preference toward 10 minutes for two reasons: hysteresis (if it ain't broke, don't fix it) and it gives me enough time to broadcast the moves by hand to John Tromp so he can comment on the game; I couldn't do this twice as fast. :) I also like the argument that some innovative ideas might just not work very well in fast games, so it would be better to keep them longer. I don't like the idea of giving extra time every move. The effect is very similar to adding a fixed amount of time, since go games have fairly constant lengths. Lags are probably tiny these days for most people anyway. And you do get an update of how much time you have left, so if you do something simple like spending a constant fraction of the time left each turn, your program will adjust itself to the lag and not run out of time. In short, I don't see what problem this solves. Álvaro. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re:[computer-go] Time Control for the new CGOS
On Tue, 2007-03-27 at 20:49 -0400, Álvaro Begué wrote: I don't like the idea of giving extra time every move. The effect is very similar to adding a fixed amount of time, since go games have fairly constant lengths. Lags are probably tiny these days for most people anyway. And you do get an update of how much time you have left, so if you do something simple like spending a constant fraction of the time left each turn, your program will adjust itself to the lag and not run out of time. In short, I don't see what problem this solves. There is a fixed amount of overhead that is being charged to the programs unfairly even under the very best of conditions, and I'm just trying to compensate for that. The idea is to guarantee that the time actually spent computing the move is an upper bound on the time the server believes was spent. Without some small fudge this is clearly impossible. Based on the times being reported on CGOS, I believe that factor should be much smaller than the 1 second I was adding. So I decreased to 1/4 second. I have also had two separate reports of programs losing on time despite the fact that they were playing instantly due to various factors which might include interface speed, network lag, even CGOS server overheads (although this appears to be quite small.) I don't want programs to lose that way. So a tiny fudge is subtracted from time spent as long as it doesn't drop it below zero. - Don ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/
Re: Re:[computer-go] Time Control for the new CGOS
On 3/27/07, Don Dailey [EMAIL PROTECTED] wrote: On Tue, 2007-03-27 at 20:49 -0400, Álvaro Begué wrote: Either 5 or 10 minutes per side is fine by me, with a mild preference toward 10 minutes for two reasons: hysteresis (if it ain't broke, don't fix it) and it gives me enough time to broadcast the moves by hand to John Tromp so he can comment on the game; I couldn't do this twice as fast. :) I also like the argument that some innovative ideas might just not work very well in fast games, so it would be better to keep them longer. One feature I'm working on is to be able to watch the games as they are being played. Then you won't have to broadcast them to John. Can't John just log into your machine and watch the games? Or I think it's easy for him to set up an ssh pipe and use gogui to watch the games from his computer. I was planning to implement something like what you are describing, but if we could watch the games live on cgos that would be awesome! Álvaro. ___ computer-go mailing list computer-go@computer-go.org http://www.computer-go.org/mailman/listinfo/computer-go/