I think most people have serious misconceptions about the effect of time on ratings. With humans, there is little effect because if you speed up the time control it affect both players (more or less) equally. I believe there are some players who will excel at fast (or slow) time controls but for the most part the effect is small.
With computer vs human there IS a big difference and it's in favor of computers playing fast. I don't think the reason is what people think it is. When a human plays faster he blunders more and has to deal with additional stress and has less time to think (duhh!) and thus plays weaker as you play faster. However, at long time controls a human still blunders, still has to deal with stress and still could do better with more time. In other words I believe the rating curve is smooth and there are no new factors introduced - reduce the time a little and you will play a little weaker. Reduce it a lot and you will play a lot weaker. Of course I know that at hyper speed you can run into hand eye coordination issues and biological limitations. However, computers blunder too and computers play weaker when given less time to think. That is no different from humans. There is also something similar to the hand eye coordination issue - a computer cannot make a move in zero time either - there is a threshold here although compared to humans it's incredibly fast - unless a physical interface is involved. Depending on the algorithm, a computer needs a little time to compute the equivalent of a 1 ply search - so there are still levels that are so fast that even though a computer might be able to produce a move, it may not be able to satisfy some minimum level of acceptability and moves produced would be blunders. Like a human the computer might say, "that's not fair, you didn't hardly give me enough time to think." I've watched humans compete at hyper blitz chess and they can play incredibly fast. Entire games with 30 seconds on the clock for instance and the moves are higher quality that (substantially) weaker players would make at tournament time controls. And it was not like every other move was a blunder either. A human can play very fast and most games lost on time forfeit are caused because the player is in a lost position anyway - or else he just has not adapted to the faster time control. So if you make a human play a game at a time control much faster than he is used to, he will lose a few games on time before "getting the hang of it." Having said all of this it's still the case that faster play favors computers - this is not disputed in computer chess and I think most people on this forum appreciate that this applies to Go also, hence this discussion. So when it comes to getting official ratings, the most critical factor is that the time control is standardized. There is one fairly simple thing that might improve this situation - once the time control is decided the human could be given "biological compensation", an extra bit of time added to each move that is a constant regarless of time control - such as 1 second per move or something like that. It's not perfect but if the right value is chosen it would reduce the error enormously between testing within a reasonable range of time controls. In computer chess it was common to play speed chess against humans - and often the computer was given and extra 1 or 2 minutes per game to compensate for move overhead - because the game was played on a physical board. I always though that was wrong if you really expected an even match. The computer should never have been given extra time for this because the human players are not given extra time to compensate them for the time delta between when they decide what to play and they physically execute the move and press the clock. In speed chess, and I'm sure in Go too, at very fast time controls the human often knows what move he is going to play even before the opponent places his stone - but it still takes time to physically execute the move. If it were not for this, a human could easily play a game in just seconds. Of course it's harder if your opponent could respond equally fast but at these time controls a human is using the opponents time very effectively, even changing your mind in the instant it takes the opponent to make his move. Don On Wed, Aug 10, 2011 at 10:10 AM, terry mcintyre <[email protected]>wrote: > Ratings can be abused, as Don has mentioned, but a little truth in > advertising should ameliorate that. > > The AGA is selling a service - its ratings - and can attach reasonable > preconditions to the use of those ratings. > > One of those would be a requirement that the conditions for that rating be > spelled out. > > "This rating was accomplished on a quadcore i7" means something different > from "this rating was accomplished on a supercomputer with 3600 cores." > > I suspect that the KGS ratings aren't too shabby; numerous high-dan players > line up in queues for a chance to test their mettle against zen and pachi. > > To determine whether time limits are "reasonable" or "too quick for > humans", look at the number of wins on time - especially if the humans were > clearly ahead when their clocks ran out. If many such wins occur, consider > adding a little more time. > > Be careful of the effect of scoring such wins on the algorithms used by the > bot. If your program stresses wins on time, you'll find it harder to win > under tournament conditions, where humans have more time. > > > > _______________________________________________ > 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
