[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [gnugo-devel] Reward multi-purpose moves ?

From: Portela Fernand
Subject: RE: [gnugo-devel] Reward multi-purpose moves ?
Date: Sun, 10 Nov 2002 04:12:12 +0100

Arend wrote:

> [patch multiplies move value by 1.003^(number of move reasons) if this
> number is at least 3]

I noticed a couple of times situations like this : the 2 top moves are very
closely valued and one of them has clearly more move reasons than the other
one. The idea with this simple multiplication is to give an advantage to
multipurpose moves in such cases (only)

> Interesting. It could just mean that connect/disconnect move reasons are
> undervalued. Since Gunnar changed these to refer to worms not dragons,
> a single move can really have a big bunch of these.

Interesting. I just made a small and stupid test : boost connection_value()
output by 10%. Surprisingly, it didn't break many things (5 passes and 3
fails for the record) and while there are a couple similarities (1 pass
and 1 fail in common), it doesn't seem to be directly related.

So I think that rewarding multipurpose moves is still an idea, which needs
some work of course. For instance, it should be qualitatively distinguished
I guess, an owl-attack is somehow more important that a strategic defense,
such kind of things...

But it also seems to me that you had a great insight, Arend. I took a more
serious look at the connection_value() function, and it seems to me that
something is indeed missing there : follow-ups. For instance, in case of
a critical dragon, the failure to connect it to a friendly living one, not
only costs proportionally to its effective size (if it dies), but it also
connects all the surrounding opponent dragons together. Shouldn't we take
this into account ?


reply via email to

[Prev in Thread] Current Thread [Next in Thread]