Re: [computer-go] Pattern matching - rectification & update

["Frank de Groot" <frank@xxxxxxxxxxxxx>]

> Impressive results.

Thanks.
Two years of hard work and a lot of almost-desperate setbacks and
uncertainties.

> Your pro-predition is sure to make a large jump
> when you include tactical status as a pattern attribute.  The only
> problem is that harvesting (or whatever your terminology) is going to
> start taking a really long time since each position has to be
> analysed.

I intend to freeze development of the pattern system at this point.
Hehe.. would you believe it, I have deleted the source for the harvesting
stage, and removed the harddisk from my PC (not really because I am
paranoid, it crashed a few days ago :)


> Do you have any plans for an evaluation function so that
> you can use these selected moves in search?

Yes, my idea is to extract a sub-database of only the smallest, most
frequent patterns and use them to sort plausible moves.
I would use a very small subset, perhaps 16,384 or thereabouts.

I have an idea for tactical search.

My problem is, I have no formal education to speak of and I am unable to
understand anything with mathematics in it.
I am therefore completely unable to understand all sophisticated
algorithms/methods for search like Thomas uses.
I simply don't understand anything of the publications.

Now, the situation is that I am therefore forced to design my own tactical
Go search algorithm.
I have been working on that for a little while (of course after reading the
literature to see what other systems are capable of, without understanding
how it works).

I can to the conclusion that even the best search methods are bad, for Go.
The parts I do understand from search algo's used in tactical search in Go
seem very illogical and inefficient to me.
The entire premisse upon which such algorithms are built appear wrong to me.
But that can very well be because I do not really understand the details.
I do not see how current search methods can be very helpful in solving
complex mid-game tactical Go situations.

I look at it this way. People are all designing a better gun to kill an
elephant with cotton balls.
All the time there are improvements.
"I made a triple-barreled hypersonic cotton ball gun that can kill a
mid-sided elephant with just a million shots in half an hour" :)

And all technological innovations lie in faster cotton balls and more
barrels, whereas real progress would be: Using bullets.

So my goal is to emulate quite accurately how a human solves such
situations.
I think its silly to go though millions of nodes.
I am already pretty sure that it is possible to achieve this. Not just by
intuition, but by design of alternative algo's on paper and calculating them
through.

I have always known this to be possible.
It is one single, simple algorithm that is waiting to be discovered.

_______________________________________________
computer-go mailing list
computer-go@xxxxxxxxxxxxxxxxx
http://www.computer-go.org/mailman/listinfo/computer-go/