[computer-go] CGOS Monte Carlo Simulations

[Don Dailey]

The K factor probably decreases too quickly.  It stops decreasing at
about 4.0.   How much it decreases as well as how a particular game is
rated depends on the opponents K-factor.   So if you lose a game to a
player who has only played a few games, it will have almost no effect on
your K-factor or in how much it decreases.

In the new CGOS, which I am working on, it will increase at a slower
rate.  

By the way - I have 4 versions of GenericMC :

  GenericMC_10000  - plays exactly 10,000 simulations
  GenericMC_100K   - plays exactly 100,000 simulations
  GenericMC_200K   - ignore - it timed out a lot and I tinkered too much
  GenericMC_VAR    - Starts with 300,000 simulations, but decreases this
                     for each move (after about 12 of it's moves, it is 
                     down to 150,000.)


The idea is to determine where the point of diminishing returns is.  It
appears to be between 1500-1600 somewhere but it's not clear.


- Don



On Fri, 2006-09-08 at 11:00 -0700, Christoph Birk wrote:
> On Thu, 7 Sep 2006, Don Dailey wrote:
> > A few days ago I posted a Generic Monte Carlo program on CGOS in order
> > to get a baseline on what to expect from a very simple Monte Carlo
> > players.
> 
> I posted a similar program (myCtest) to see if it gains a similar
> ELO rating. The first version had a couple of bugs but instead
> of running the fixed version under a different name I just kept
> the old name ...
> The problem now is that CGOS' rating algorithm is not appropriate
> to handle players with improving strength.
> The K-factor appears to decrease monotonically (fairly quickly)
> to a low value (4?).
> How about making it "dynamical" by allowing the K-factor to increase
> when  a group of unlikely game results show up; e.g. a program wins more 
> than it's "fair" share of games against stronger opponents.
> 
> Thanks for CGOS, it's a great resource!
> Christoph