Re: [gnugo-devel] Dragon safety early in the game

[Arend Bayer]

On Wed, 28 Aug 2002 address@hidden wrote:

>
> > Of course dragon2.weakness was introduced exactly to address this
> > problem. I still think that it is a useful way to go, but it definitely
> > needs further tuning. I know I have to add documentation, and it has to
> > go in the -w output, too.
>
> Regarding adding documentation, could you at least e-mail some
> more explanation to the list?
>
> I think that the early game dragon safety issue is an area where
> we could get some immediate improvement in strength.

Agreed. I have already explained how it gets used in my previous e-mail.
I should point out it is not yet used for connect/cutting moves.

As for how it is computed: There are sufficiently many comments in and above
compute_dragon_weakness_value() for details, so I'll restrain myself to
a short summary.

The 3 main inputs are:
1. territorial moyo value (post owl analysis) (in fact the most imporant
        ingredient)
2. escape value
3. number of eyes according to current estimate.

>From each of these value, a weakness value betweeon 0.0 and 1.0 is
computed, using gg_interpolate with the curves def'ed above the function.
(E.g. 0.3 for a territory value of 6 resp. an escape value of 10, or
0.05 for a territory value of 12 or an escape value of 20).

Then the weakness computed as the mininum of these 3 values (with small
correction coming from the 2nd best value).

[Note: In the tuning I have done I concluded that no. 1 is the most useful
input. This is indirectly taken into account -- in most of the cases,
the weakness computed from the territorial value will be the lowest of
the three, hence the one with the most impact.]

These value then gets corrected up by at most 0.125 resp. 0.15 for
uncertain owl attack, uncertain owl defense resp. existence of an owl
attack threat.

Hope this clarifies a bit.

Arend