[next] [prev] [up] Date: Mon, 16 Oct 95 23:08:02 -0400
[next] [prev] [up] From: Jerry Bryan <BRYAN@wvnvm.wvnet.edu >
~~~ ~~~ [up] Subject: Correctness of Large Searches

When I was in graduate school in another life many years ago, one of
the big issues was program correctness, and whether programs could
be proven to be correct. As I recall, the short answer is that
non-trivial programs cannot be proven to be correct. I haven't
really followed the issue since then, but I doubt that the answer
has changed.

Also, many mathematicians are suspicious or unaccepting of proofs or
other results which involve large computer searches. For example,
the proof of the Four Color Map Theorem is a famous result involving
a large search which is held in certain suspicion. For one thing,
the referees for the paper were not able to reproduce the results
of the computer search (not enough computing power). For another thing,
the referees were neither able to confirm that the programs were error
free (who can?), nor that some sort of computer error (software,
hardware, or procedural) did not occur during the running of the
programs (c.f., the famous Pentium floating point divide error).

Two factors cause me to raise the question of program correctness
at this time. One is simply that some of the searches we do are
so large, how do we know that the answers are correct? The other
is that I have found an error in one of my (smaller) searches that
I need to report.

Those of you who actually fight through the details of my longer, more
boring posts will recall that there are essentially three different
models I use from time to time. For cubes without centers, I usually
use representative elements of sets of the form {m'Xmc}. For cubes
with centers, I use either representative elements of sets of the
form {m'Xm}, or else I use representatives of the form
Repr{m'Xmc}*C to simulate Repr{m'Xm}. The latter model is very
compact for cases where complete searches can be accomplished

For the case of the 3x3x3 cube, corners only, with centers, q-turns
only, I recently discovered the following anomaly. Note in
particular level 8 and below of the search.

3x3x3 Corners Only -- Qturns

Distance    Repr{m'Xmc}*C        Repr{m'Xm}
    from          Model             Model
 0               1                 1
 1               1                 1
 2               5                 5
 3              24                24
 4             149               149
 5             850               850
 6            4257              4257
 7           16937             16937
 8           57800             57848
 9          180639            180787
10          466052            466220
11          676790            676786
12          392558            392342
13           45744             45600
14             163               163
Total        1841970           1841970

The 3x3x3 corners only case has been searched a number of times by
a number of people, all of whom got the same results. However,
these results all were for every position, including those which are
M-conjugate. My searches are for conjugacy classes only, and therefore
I have nobody with whom I can compare results directly. The only
way I can compare results is to expand the conjugacy classes, and
regrettably I did not do so when I first calculated the corners
only case.

For a variety of reasons to be detailed below, I believe the {m'Xm}
results above are correct and that the {m'Xmc}*C results are incorrect.
But at the same time, I do not believe there is an error per se in
the {m'Xmc}*C model. Let's see if we can make some sense of this.

When I first discovered the anomaly, my reaction was that
the {m'Xm} model is simpler, and Occum's Razor suggested that
the error was in the {m'Xmc}*C model. Also, when I expanded
the conjugacy classes for the Repr{m'Xm} model, the results
matched what everybody else had posted.

Hence, I went back to the
{m'Xmc}*C program, which I hadn't looked at in years. After many
hours of reviewing the model, and many more hours of reviewing
the program itself, I not unsurprisingly found no errors.
Furthermore, I was no longer convinced that Occum's Razor still
applied. The {m'Xmc}*C model is "more complicated" in some ways,
but on the other hand, every cell of storage for the data base
can be addressed directly. There is no sorting, merging, and
matching of huge files as there is with the {m'Xm} model.

So I ran the {m'Xmc}*C program again. Very surprisingly, this time
it produced different answers, and the answers were the same as for
the {m'Xm} model! So what's going on? I don't know.

I am presently a bureaucrat (cubing is a hobby), but in previous
lives I have been a technical support person. In that role, I
have found numerous problems that caused programs to produce incorrect
results, even though the program was "correct" (whatever that means).
I have found hardware errors vaguely similar to the Pentium error,
operating system errors, compiler errors (especially with optimizing
compilers), and subroutine library errors. So my best guess is that
something of this nature caused the {m'Xmc}*C program to produce
incorrect results one time and correct results another time.
Of course, an uninitialized variable or pointer can cause similar
symptoms, but I have not been able to find anything like that in
my program.

The program is written in Turbo-Pascal for a PC using DOS. I have
the exact same compiler now I have had for ten years or so. The
Pascal source code is unchanged from when I ran it before. The
first time I ran the program, I ran it under DOS, or maybe
in the DOS box in an early version of OS/2 (can't remember for sure),
and it ran on a 286 or an early 386 (can't remember for sure).
This time, it ran in the DOS box under OS/2 Warp on a 486/50.
So lots of things have changed. Furthermore, both times I ran it,
I used the VDISK facility to cache all the disk files in memory,
and OS/2 Warp surely has a newer VDISK than whatever I was using
before. I even wondered if the data base was correct the first time,
but maybe I had a bad counting program analyzing the data base. But
I still had the original data base and counted it again. It
really was wrong. So the mystery of the incorrect results is not solved.

In light of the above discussion, I thought it might be appropriate
to summarize some background about my larger searches. I want to
indicate which ones of them seem pretty well verified, and which ones
of them might benefit from further study.

<U,R>           -  I believe this one is ok.  I ran the q turn case
                   with and without conjugacy classes.  Upon
                   expanding the conjugacy classes, the results
                   matched the results without conjugacy classes.
                   Also, the results matched results posted by
                   Mark Longridge as far as they went (although
                   my anomaly with corners-only suggests that
                   such matching doesn't prove very much).  I
                   ran the q+h turn case only with conjugacy classes,
                   and expanded then to get the results without
                   conjugacy classes.  The model was Repr{w'Xw}
                   (W-conjugacy instead of M-conjugacy), with
                   much sorting, merging, and matching of external
 edges only          This ran for about a year, so there is potential
(without face        for error.  The model was Repr{m'Xmc}.  The
  centers)           runs to create neighbors were performed primarily
                     on two 486's and a 386, using a mainframe as
                     a file server.  A UNIX station was added for
                     the last few months.  All sorts, merges, and
                     matches were run on the mainframe.  The whole
                     process was driven by REXX scripts on the PC's
                     and UNIX stations (we run REXX on all our UNIX
                     boxes).  There were actually more lines of code
                     for the scripts than for the programs.  The
                     scripts moved files back and forth with FTP,
                     and contained code to detect and correct errors
                     with the FTP's.  The problem has only been
                     run once, so there is no independent verification.
 edges only           This was run twice, once using the {m'Xm} model,
(with face            and once using the Repr{m'Xm}*C model.  The
  centers)            answers matched through level 11.  They did not
                      match at level 12.  I have only posted through
                      level 11 to the list.  I have not had time to
                      find the error.  I do not expect to find the
                      error in any of the programs.  It took about a
                      year to get to level 11, running on the mainframe.
                      I expect the error to be somewhere in a script,
                      probably failing to recover from some error
                      condition like a system failure.  These things
                      run around the clock, and inevitably get caught
                      in system failures from time to time.

Although I was not able to do so, this is probably
the largest cubing problem where we can conceive
of running the problem to completion using today's
technology. Had I been able to do run the problem
to completion, the Repr(m'Xm)*C model would have
served as verification for the "without face
centers" case because the minimum of each
row of the solution matrix with centers serves
as the solution for the without centers case.

Whole cube            These are the most important runs.  The problem
(edges and corners    has been run twice using the {m'Xm} model, although
 with face centers)   you might not wish to count it as twice.  The
                      difference is that the edges were the major sort
                      in one run, and the corners were the major sort
                      in the other.  The two runs did match, which is
                      a good (but not perfect indication) that there
                      was no error (e.g., such as the
                      scripts failing to include something in
                      in a sort or merge that should have been
                      included).  It would still be nice to have
                      verification by someone else.  I was able to run
                      up through level 11.  I did not try level 12
                      because it was just too big.

 = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = =
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

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