Comments/Ratings for a Single Item

SMIRF still is an 8x8 and 10x8 multivariant engine and GUI at amateur level. Experiencing some problems when being executed at Windows XP 64 Bit I recompiled an actual version of SMIRF and made it downloadable from http://www.chessbox.de/Compu/schachsmirf_e.html. Thus SMIRF still will have its conceptual weaknesses during mating operations. Nevertheless some 10x8 fans might welcome the current version made available. Please report on still existing weaknesses of the GUI or the download link. SMIRF itself will not be improved, instead there will be a rewritten private engine Octopus.

Indeed,this is a known problem with Smirf. Because the underlying piece-value model is linear in the average piece mobility, the piece values become additive, and a the value of Q becomes that of R+B, and that of A that of B+N. For A this is more disastrously wrong than for Q, but trading Q for R+B is still quite bad (like blundering away a Pawn).

For the short-range leapers I found a clear non-linearity in the relation between (maximum) number of target squares N and piece value V:

V = (30+5/8*N)*N (centiPawn).

The methodology of basing piece values on board-averaged mobilities seems flawed to me: it overestimates the impact of bad squares where the moblity is low. In practical play you avoid putting the piece on such squares. e.g. take a few thousand positions randomly chosen from grandmaster games, and count how many of those had a Knight on a corner square. It seems a safe bet that this will be FAR LESS than 4/64 = 6%, and in fact I would be really surprised if it is more than 0.6%.

It would be interesting to observe the frequencies with which pieces visit each board square in grandmaster games, and determine how this correlates with the mobility of the piece.

I appreciate the 3 versions of SMIRF loaded with different CRC material
values that you sent me for testing purposes.  I realize compiling them
was not a productive use of your time toward developing Octopus or
creating future versions of SMIRF.  So, I sincerely hate to complain.

Internal Playtesting- Scharnagl
http://www.symmetryperfect.com/pass

Push the 'download now' button.

I played one game of Embassy Chess (mirror) at 40 minutes per move.  The
white player was version 0 (standard) and the black player was version 2
(highest archbishop value).  The black player won.  However, the victory
was not attributable to the white player valuing its archbishop too low in
an exchange.  Instead, it was attributable to the white player valuing its
queen too low in an exchange.

White traded its 1 queen for 1 knight + 1 rook belonging to black.  This
gave black a 3:2 advantage in supreme pieces which, over the course of the
game, was reduced to a 1:0 advantage in supreme pieces which gave black the
ability to out-position white in the endgame, gain material and win.
The game was not even close or long ... ending in 53 moves.  I have seen
this happen many times before.  Of course, with version 0 and version 2
having identical material values for the queen, rook and knight, it could
have just an likely 'thrown the game away' to the other player.  That is
the reason I cannot continue playtesting with what you provided to me.

Under the Nalls model (for example), there are 3 supreme piece(s)
enhancements:  the non color-bound enhancement, the non color-changed
enhancement and the compound enhancement.  In CRC, they total a 43.75%
bonus for the archbishop above the material value of its components (the
bishop and the knight), a 12.50% bonus for the chancellor above the
material value of its components (the rook and the knight) and a 18.75%
bonus for the queen above the material value of its components (the rook
and the bishop).  The entire purpose of the supreme piece(s) enhancements
is to provide a measurably appropriate deterrent to trading any supreme
pieces too lightly to your opponent thereby ending-up with a potentially
game-losing disadvantage in the ratio of supreme pieces.  The Muller model
is similar in this respect.

If I had to choose only ONE foundation, experimental or theoretical, for
my model, then I would choose experimental without apprehension.  Of
course, I am allowed to use both.  So, I do because I remain hopeful that
eventually, thru relentless effort, my theory will attain a worthwhile
condition (that has previously eluded it) whereby the theoretical and
experimental foundations will become mutually reinforcing.

I would characterize my position as regarding both the experimental and
theoretical foundations as important (although I definitely consider the
experimental foundation primary).

I would characterize Muller's position as being that the experimental
foundation is everything that matters and the theoretical foundation is
just an unneeded crude, inaccurate approximation to experimental numbers
decorated with arbitrary words and concepts.  Maybe so?

I would characterize Scharnagl's position as being that the theoretical
foundation is supremely important as it must dictate and predict the
optimum experimental numbers.  [I agree that a great theory should be
expected to do so.]  Furthermore, the theory must be elegantly simple and
intuitively accessible.  [I consider this expectation unrealistic and
impossible.  Generally, the optimum material values for chess variants are
too complex in their estimation-calculation to be reducible to simple
formulae without sacrificing accuracy to an unacceptable extent.]

Scharnagl:

Please reconsider revising your CRC model even if doing so unavoidably
renders your theory somewhat more complicated in its concepts and
formulae?  The playing strength of SMIRF (standard version) can probably
be improved significantly by taking such steps.

Well, Sam, I am not that rich, so it would help somehow to have some minor contributions e.g. for to buy some more recent development tools and for to update my hardware from time to time. But it is more disturbing to experience people seeing no value in such software as long it is free. Thus freeware is devaluating the scene and does a bull's job to chess variant software. Making the entry version of SMIRF a donationware should have been signaling that message, but most of SMIRF's users nevertheless mixed that kind of releasing up with freeware. 

Another demotivating detail is, that though I have tried to publish all of my chess pages using both German and English languages, no (as far as I know) native English speaking chess page had been turned to become bilingual too by also supporting German, that way following my example. This makes me feel like a second class person. Therefore I have decided to no longer continue such an approach, instead I will merely use my German language, if I ever would relaunch my web site again. 

Thus developing Octopus out of SMIRF's bones is done rather slowly and privately. That would not exclude really interested people from contacting me or following the project, nevertheless I am ignoring the public, e.g. by chosing Mac OS X as development system, saying 'sorry' to MS Vista, and by redirecting my old site to a small German language blogging site. Reinhard.

Don't feel bad. I have worked hard making my own open source project. Have I gotten a single cent for this project? No.

- Sam

The link to SMIRF is dead.  A transition is underway that will take a long time from SMIRF for IBM-compatible computers running MS Windows to Octopus for Mac computers and OS.

SMIRF > Octopus
(in German)
http://web.mac.com/rescharn/iWeb/Octopus/Blog/Blog.html

The last version of SMIRF released was BC-168a.  If you want it but missed your chance to download it, please send me an E-mail privately.  I can send you the file.  Please remember it is donationware- donations to its developer, Reinhard Scharnagl, are encouraged!

Hi Mats, let me explain, that my understanding of AI is NOT simulating
anything but developing solutions genuine to machine. Intelligent behavior
is not merely imitating. Now, how to support the SMIRF project:

a) intellectually: by showing interest, writing comments, testing, giving
constructive critic, mentioning the SMIRF project where it would make
sense. There is no need to exceed intellectual limits, we all mostly are
specialists, thus don't worry when there will be gaps, I don't either.

b) financially: where fans like to use the SMIRF program, they should be
aware, that there are costs for organizing and maintaining hard and
software, e.g. actual development environments, and that it needs a lot of
time, wherein no money could be earned. Thus it makes sense to those people
to place some donations via paypal.

Hi, Mats, beside of the fact, that the Windows distribution of SMIRF still
is the only one and just has been updated (in the bonus version), it does
not matter to me, how many people are using Smirf, but in contrast it
does, how many are really supporting it. And, because of this number is
nearly vanishing, I simply decided do migrate to Mac OS X because of
software political reasons.

Yesterday I have published on my web site my new model for calculating the
average piece exchange values for various gaits, both: simplified and
improved, at http://www.chessbox.de/Compu/schachansatz1_e.html. Reinhard.

Hi Mats, one of my first goals is to make SMIRF play self-containedly
especially through the opening stage. I am happy, that this works even for
randomized starting positions half-way sufficiently. To manipulate the
opening behaviour to gain more success simply by adding some quirks
instead of modifying the global model concept will destroy the chance of
finding an abstract and consistent base model. Still I am e.g. not fully
satisfied with my average piece value model, and the evaluation function
is three times as slow as necessary.

Nevertheless I have thoughts to distribute a predefined amount of
evaluation to be randomizedly disposed in chosing second class moves
during the beginning to gain a less determinated play at fixed opening
arrays. Actually I still recommend to play Chess960 or 10x8 CRC to have
more diversified game experiences.

The future of SMIRF: well, in the meantime there have been some questions
from different sides, what will be with SMIRF. 

First, I am about to migrate it to the Mac OS X. That is, because I am
angry about MS Vista and its unsufficient driver set (especially 64 Bit),
its bad behaviour against programmers and its several backdoors for
government, police or secret service investigations targeting every file
on a Vista system (as I prosume). In contrast Mac OS X is again going to
create a homegeneous 64 Bit user landscape starting with coming OS X
Leopard version.

SMIRF will become 64 Bit then. There will be no big speed step by that,
because SMIRF is no a bitboard based program. But it will use the grown
register set, and it will then benefit from any compiler optimization
improvement. I remember days, where I investigated to have an ARM RISC
processor based solution, but there are currently no fast desktop systems
or sponsors, which would provide me with such a development environment.
Thus I stay on Intel/AMD64 based hardware (unfortunately also without any
sponsor yet).

SMIRF will not be merely migrated to Mac OS X. I intend to have a lot of
new design elements:

a) redesign the internal data structure and caching logic
b) rewrite the position detail evaluation function (which is still the
first slow draft, may be a 200% speed up then)
c) support additionally also an additional piece (General = N+B+R)
d) support blocked squares (to customize the playable square zone)
e) build in prompting (instead of badly working permanent brain)
f) attempt to implement multithreaded multiprocessor usage 
g) create a separated unit for to contain move logic and pgn file
maintenance beside of the GUI
h) attempt to have a UCI protocol based GUI

That is a huge project and will need a remarkable amount of my free time.
So do not wait for a finished Mac OS X version in the near future. May be
some of then made progresses will be reimplemented into the working
Windows 32 Bit solution, but from now on possible updates would only be
published for the donators' bonus version.

Reinhard.

Smirf clearly improves its positional play when given more time. This is
unusual. But it has a rather passive style of play. Another thought:
different alternative variants can also be achieved by keeping the same
pieces but introduce the Gustavian board (
http://hem.passagen.se/melki9/gustaviii.htm ). S Trenholme did this in
Capablanca Gustavian (zrf) which can be downloaded from yahoo (chess
variants). It is an interesting concept. The smaller board might affect
the unruly Archbishops and Chancellors so that they are somewhat
neutralized.
/Mats

To uncover the *laws* underlying any subject matter is the gist of the
scientific paradigm. I am surprised  that there are people who think
differently, which is interesting, of course. Another thing: there are so
many chess variants on this site that are better than those Capablanca
variants, with their rather brutal pieces. Personally I even prefer the
Amazon to the Archbishop and Chancellor. The Amazon is easier to handle,
and it must hide to all other pieces, so the games are easier to predict.
/Mats

Hi Mats, please have a look on that, what you have said. You are using such
names like 'central squares' or 'established chess laws'. But that kind
of arguing is just the problem. Why are central squares that important?
Where is the center in different chess960 games, where the king himself
might be decentralized? By that you learn, that such categories are
derived from more essential but abstract things. 

Another more transparent example are the average piece exchange values:
where do those traditional values come from? What is with that at
different variants? At my webpage you will find a simple theory to derive
such values even for unusual piece types or board sizes. As I tried to
tell you, I am not interested in that kind of chess knowledge, where it at
least leads to a simple implementing of a copycat behaviour. 

If you want to create an effective and INTELLIGENT chess program, you
first have to UNDERSTAND the basics of chess, not to imitate the so called
chess knowledge, which is moreover differing enormously depending on which
chess master you will ask. Then you will have to TRANSLATE it into the
world and language of a CPU.

Today there are a lot of effective but mostly huge chess programs. So
there is no urgent need for to write another one. But on the other side
there are very few intelligently working approaches using instead very
restricted means. I am arguing for to have computer chess program
tournaments with LIMITED means, especially targeting the persistant storage
size including the program executeable. They should be bound to a special
upper bound. And it should be measured in a packed form e.g. as RAR for to
skip inner redundancies as generated depending of the selected computer
language and to avoid the temptation to undergo any limits by including
packed knowledge.

Of course, I did not mean 'stupid looking up', like opening books and
endgame tables. I did not mean concrete knowledge. What I had in view are 
the established *chess laws*. For instance, in the opening you
must direct attention to the centre. There are two methods, either a
direct fight for the central squares, or an initial forfeiture followed by
an immediate undermining of the points of support. Flank operations must
not begin before the situation in the centre is clarified. In the endgame
the king must become active, and take heed of opposition, etc., etc. In my
own weak little DOS program ( http://hem.passagen.se/melki9/blindc.htm )
there is no book knowledge, either. But it tries to control the centre,
nevertheless. And it seems to play the openings rather well, without
opening book. 

Facts are that programmers are reluctant to teach the programs this kind
of knowledge. Instead they want to create as effective algorithms as
possible, so that the correct move is reached anyway. This creates a form
of chess that is lacking in variance. There are very many ways of handling
a position, provided that you follow the chess laws. If you don't follow
the chess laws, but only calculate, then the program will decide for only
one possibility. This is a faulty conclusion while there are other moves
that are just as good.

In this sense, I'm afraid, this project is similar to other chess
software solutions in that you put to much trust in the calculative
capability of the program. I don't see why abstract knowledge cannot be
combined with an AI approach.
/Mats

Please let me explain another way ...

The purpose of SMIRF is to fairly play ALL 960 FRC starting positions and
ALL 12,118 CRC starting positions equally and extraordinarily well.

Obviously, it would be impossible to generate high-quality opening books
of adequate depth and width for all of these 1000's of games within a
survivable time.  So, to attempt such a project is not even being
realistically considered.

Moreover, some people are skeptical that any of the starting positions
with FRC pieces upon the 8x8 board or CRC pieces upon the 10x8 board are
at all or significantly superior compared to other unknown permutations. 
In other words, they do not consider Chess any better than the other 959
FRC variants nor do they consider any of the 2 dozen CRC variants given
names by their inventors any better than the other 12,000+ CRC variants.

[Personally, I think otherwise that Chess & Opti Chess are the best FRC &
CRC games, respectively.  In the latter case, I admittedly lack
impersonal objectivity.]

In my playtesting experience, SMIRF is vastly better than merely 'an
impractical experiment that plays Chess poorly' as you describe it (more
or less).  You are missing the point.

You obviously do not fully understand that there is a strong correlation
between the quality of moves generated by a pure search intelligence
program (such as SMIRF) and the time, plies or positions that must be
invested to achieve worthwhile results.  When testing one program against
another or testing one set of material values for pieces against another
using the same program, I NEVER use a time per move of less than 1 hour
... running a dual-CPU (2.4 Ghz per CPU) server.  I regard
quickly-obtained results as random, too replete with bad moves and
analytically uninteresting (since a sub-genius such as I could have won
the game playing either side).  Given reasonable conditions, SMIRF &
ChessV make moves well beyond my capacity as a player.  By comparison,
the Zillions Of Games program makes bad moves even when given 24 hours per
move or more.