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

Re: [computer-go] AI decision making systems



----- Original Message ----- 
From: "Peter Drake" <drake@xxxxxxxxxxxxxxxxx>
Subject: Re: [computer-go] AI decision making systems


> Is MogoGo a program that can play on its own, or just a study aide?

It can play on its own, since last week.
But to score positions I use GNU Go.


> > This only works when millions of patterns are stored that have
> > "cross-correlated" values. Those patterns again are used to reduce the
> > input space for a neural network.


> Do your patterns include "don't care" positions?

No.
I found that when you include enough "do care" patterns, all the other
patterns automatically become "don't care" by default.

I have to stress that this is only possible when one includes a sliding
scale of very small to very large patterns, and includes "enough" patterns
of the smallest varieties.

The possible number of normalized patterns in my smallest pattern:

  x
 xxx
xxxxx
 xxx
  x

is 3^12 / 16 = 33215 patterns. Indeed my pattern system includes all, but it
includes many more variations on those, namely stone count ranges and libery
count ranges of adjacent chains, which explodes the number of patterns just
for this smallest pattern size into the low millions.

Only with such a comprehensive system can one "safely" ignore "don't care".
Effectively, all that is a realistic pattern to play on is already included.
This by simply recording all of these patterns from 400,000 games x 250
moves = max. 100 million variations. With such a large sample size, every
possible pattern in the smallest size class occurs several times, which
allows us to "say something" about this pattern in relation to the other
patetterns.


> Do you have a paper explaining this system in more detail than the page
> above?

Sorry, no. I have no higher education so I am not familiar with writing
papers. I am currently studying on how to write a proper paper. Would it be
acceptable to write a paper explaining the general architecture & obtained
results whilst still not disclosing technical implementation details?

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