[computer-go] CGOS Monte Carlo Simulations

[Magnus Persson]

Quoting Don Dailey <drd at mit.edu>:
>> Having the possibility to start the rating from an "expected rating" 
>> would be
>> good. So that, the rating would converge more quickly. I don't know if we
>> must think of a way to avoid cheating like "I expect a rating of 10000 :-)".
>> Limiting the expected rating to a realistic range, and starting with a high
>> K-factor would be sufficient.
>
> The way I rate unknown players is sub-optimal.  There are more efficient
> ways to get a quicker convergence for a new player and I will probably
> implement something much better for this.

Here is a simple idea. When a new version logs in for the first time it could
inherit the rating of an old version if the name and password for the old
version is sent as extra parameters. Using the rating of the old 
version is the
best guess because most changes to a go program do not change playing strength
measurably and if there is a change it is rarely an improvement. (If strength
always improved when you changed the program most programs would be 
very strong
now).

-Magnus