On Wed, Jul 9, 2008 at 8:17 AM, sapphirebrand <[EMAIL PROTECTED]> wrote: > > --- In [email protected], "marc_roddis" <[EMAIL PROTECTED]> wrote: > > > > Perhaps I should make my question more specific. What are the > > ratings of the people that are beating quackle half of the time > > (for example against speedy player)? > > Marc > > Good question, Marc. A well-formulated attempt to get a straight > answer. But you won't get a straight answer. :-) > > My finding with Maven was that some players could defeat Maven at far > above their expectation according to the rating system. And some were > far below. > > In those days a rating of 1800 was about 300 rating points lower than > the peak human rating. Such a player should have ~10% chance (?) of > beating Maven at the top level (which was non-simming in those days). > Yet I know one 1800 player who was able to win almost 33% of her > games.
I would not be surprised if there were 1800 players who happened to have beaten a specific championship level player 33% of the time, yet beat other top players only 10% of the time. I would not be surprised if there were championship level players who happened to have lost to a specific 1800 player 33% of the time, yet lost to other 1800 players only 10% of the time. This kind of variation is more due to the nature of the game rather than the nature of programs that play the game. (This is what makes the ELO rating system much less stable/reliable in Scrabble than in Chess). Steven Gordon > > I know of two winners of multiple world or national championships > whose tournament ratings were comparable, yet one could beat Maven > 49%, and the other was around 35%. > > BTW, there are several hundred games (over a thousand in some) in > each of these samples. Maven would keep a record of every game that > you played, so I was sure that the differences were real. > > It seemed to me that the players who like to keep things wide-open > were at a disadvantage. That style probably helps them in their human > games, but it is a problem against computers. > > None of that ever bothered me, though. People wanted to know how > strong Maven's levels were, so I gave them a straight answer. I just > set the top level to 2100, and calibrated successive levels at 50 > rating points lower and lower. Based on self-play. > > This system does not guarantee comparability to human tournament > rating. Possibly the Quackle developers can devise a technique that > is comparable. Still, individuals will vary from that baseline > depending on their style. > > Best, > Sapphire Brand > > PS: Odd, when I submit my posts online, Yahoo asks "What language is > this post in?" They give me a choice of English, but not TWL or > SOWPODS. >
