[next] [prev] [up] Date: Fri, 11 Nov 94 22:06:49 -0500
[next] [prev] [up] From: Jerry Bryan <BRYAN@wvnvm.wvnet.edu >
~~~ [prev] [up] Subject: Re: <U, R> Searches
On 11/11/94 at 15:36:00 mark.longridge@canrem.com said:

With the computer hardware I'm using (486-DX40 with 4 megs) and
my current algorithm I created a file "ur.dat" which is a flat
ascii file. It contains all the processes which generate distinct
positions up to 12 q turns. I also have the file "ursum.is" which
contains all the `cubesums' for each element of <U, R>.

I have considered keeping a file of processes, either in parallel
with my file of positions, or together in the same file. However,
I have always rejected the idea because it would take too much
storage for the very large searches I want to accomplish.
Also, since I only store representative elements of equivalent
classes in my data base, it is hard to know what a "process"
means (see below).

Using this method I have found a process for all <U, R> positions
with the longest being 22 q turns (so far). The hardest positions
can take as long as a couple of hours. I have no idea what the
antipodes look like at this time, but I'll probably try the
random approach soon.

Interesting approach, and very different than what I do. I simply
do a "game theory" type tree search of positions (*not* processes)
with Start at the root of the tree. I have to search the whole
depth of the tree rather than half the depth of the tree, as you do.
(I could obviously verify the depth of one particular position with
two half-depth searches, but I confess I have never been very
interested in "particular positions". I want to search the whole

I tend to think of the processing required for whole tree searches
as "global" because you cannot determine anything about
one particular position except in the context of the other
positions down to the same level of the tree. By contrast, there
is some interesting processing that you can do that is "local";
e.g., conjugate class sizes, symmetry group determination, etc.
can be determined on a position by position basis without
regard to any other position. Such "local" processing is generally
much easier than the "global" processing. "Local" processing
is O( N ) and requires essentially no memory. "Global" processing
is O( N log(N) ) if you are efficient, maybe O(N^2) if you are not,
and is very consumptive of memory. As I have said before, I solve
the memory problem by externalizing the data to files and sorting
the files.

Saying that I have a hard time converting positions into processes
doesn't mean that I cannot do so. The process for a given position
can be extracted from a data base of positions by backtracking
from the position back to Start. However, I have two difficulties.

One is that I store only representative elements of equivalence
classes (of the form {m'Xmc} for centerless cubes, or M-conjugate
classes of the form {m'Xm} for cubes with centers)}. That means
that as you backtrack, the process you are determining (backwards)
will rotate and reflect out from under you. For example, suppose we
have Repr{m'Xm}=f'Xf for some fixed f in M. If Xq for some q in Q
is a neighbor of X which is one move closer to Start than X,
then we might have Repr{m'(Xq)m}=g'(Xq)g for some fixed g in M.
But f and g need not be, and usually aren't, the same.

It might be noted in passing that there is duality between the
symmetry of a position X and a process P=p1 p2 ... pn which produces
X. That is, if f is a fixed element of M, then we have
f'Xf = f'Pf = f'(p1)f f'(p2)f ... f'(pn)f.

The second problem is that my files are so large that I have to
keep them on magnetic tape. If I need to find a process, I first
find a sequence of positions (really, a sequence of representative
elements) on the tapes (hard to search tapes!), and copy the
positions to a small disk file. From that point, it is not
especially hard to run a program to display the positions in
sequence and determine the process.

 = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = =
Robert G. Bryan (Jerry Bryan)              (304) 293-5192
Associate Director, WVNET                  (304) 293-5540 fax
837 Chestnut Ridge Road                     BRYAN@WVNVM
Morgantown, WV 26505                        BRYAN@WVNVM.WVNET.EDU

If you don't have time to do it right today, what makes you think you are
going to have time to do it over again tomorrow?

[next] [prev] [up] [top] [help]