[Bug-gnubg] Re: Checkerplay vs cube decision errors

[kvandoel]

On Fri, 21 May 2004, Joachim Matussek wrote:

> > > Should be  self-explaing now. It  is not b(N)  which goes down.  It is
> > > b(N)/(No. of close or actual cube decisions) which goes down.

> > I don't get  it. If you reduce  the cutoff "No. of close  or actual cube
> > decisions" goes down. b(N) also goes down so the ratio remains the same.

> you seem to be right.

> It seems  that (a2(N)/(No. of unforced moves))/(b(N)/(No.  of close or
>actual cube  decisions)) only  depends on the  match length.  It ranges
>from 3.38 to 2.56 within the chosen examples.

> Thus we  should be able to  conclude that checkerplay  errors and cube
>decision  errors have  different  weights when  calculating a  player?s
>rating. It also means that they have different influence on the outcome
>of matches.

> Yes?
> No?
> If yes, why?

> Isn?t our term and calculation  of cube decision errors flawed? Should
>we estimate cube errors in a different way?

The  idea behind my  simulations was  to just  consider the  chequer and
error rates,  however computed, as probes  into the system,  and see how
they correlate to the actual rating of players as obtained by estimating
their  playing  strengths statistically  from  the  luck adjusted  match
results.

The formula  that's currently programmed  into GNUBG to  translate error
rates into rating  point losses over a match is a  good empirical fit to
the data.

How to interpret the individual terms in the formula and if you can draw
any grand  conclusions about the  importance of chequerplay  versus cube
decisions is less clear.

However,  in practice, without  exception, the  number of  rating points
lost due to chequer error (QLO) is always much larger than the number of
rating  points lost due  to cube  errors (CLO),  in a  particular match.
(Well maybe  if someone cubes  after opponent opens  31 and then  gets a
forced game  with a closeout and no  real chequer play you  would see an
exception.) This  seems to  hold across all  ratings, though  the actual
ratio CQ = CLO/QLO is player dependent (and of course fluctuates a lot).

It seems reasonable to interpret  CQ as the balance between chequer play
skill and cube  handling skill of a  player, CQ = 1 means  the player is
equally skilled in both. In  practice, for every human I've observed, CQ
is much larger than 1, of the order of 10.

We can  conclude that  chequer play is  more important in  the following
sense.  Suppose there is a bg pill on the market. The red pill gives you
perfect chequer skills,  the green pill gives you  perfect cube handling
skills, but you can take only one  pill at a time due to disastrous drug
interactions.  Which  pill should  you take?  Clearly  the red  pill for
almost everyone.

The small  effect of cube errors  on overall performance  allso gives an
interesting  perspective on  the so-called  "pip-counting"  methods.  It
follows that for almost everyone,  learning to accurately pip-count is a
complete waste of time.


Kees