Re: Pattern learning: update

[Antti Huima <huima@xxxxxx>]

On Wed, 9 Jun 1999, Henrik Rydberg wrote:

> Your algorithm has a lot of potential, I think. Now you're
> using it to pick patterns from real games. What if one were
> to construct lots and lots of small random situations, in a
> way that they can be played, locally and using exhaustive search,
> to the "end" somehow. Then one could create patterns automatically.
> 

> * Solve the inner problem as in Tsume-Go (David Dyer, do you
>   have a comment on this?)
> * Match the outcome using Huima's algorithm.
> 
> Of course, the assumption "alive and kicking" is crucial and wrong,
> but hey, this is only a construct in order to produce move
> *suggestions*, so it might be ok.

I understand this idea, but I'm not sure if it is much more than caching
the results of the local searches. The point with learning from expert
games is that the situations where the experts tend to play are "typical"
to expert play as opposed to the random situations you perhaps seem to
propose, and so the patterns learned from the expert games can be "more
useful".

I don't claim that your idea is bad. I think it is a good idea to search
through commonly occurring situations and variants of them and store the
results so that they can be used from one game to another. This is similar
approach to that which is routinely taken for chess endgames. However, I
assume that running the search results through VQ variants does not
necessarily improve the value of the caching when compared to just storing
the results as they are. The point is, after searching you understand the
local situation well and can memoize the result and generalize by
searching through position variants. When learning from expert games, you
do not understand the situation but see nevertheless the expert moves and
need to be able to generalize without proper understanding.

These are, however, mainly feelings.

-- 
Antti Huima
SSH Communications Security Oy