🚀 go-pugleaf

In article <73@milford.UUCP> bill@milford.UUCP (bill) writes:
>
>About a week ago the 5 MHz Fidelity Challenger which I had ordered
>from the USCF arrived and I was rather disappointed with its play
>against me, ....
>
>The 5 MHz machine's faults are: tends to accept the 'bad bishop',
>accepts bad pawn structure without compensation, unable to find a
>non-destructive move in positions in which constructive moves are
>difficult to find, on the other hand it seems much better
>tactically.
>
>But a deeper problem is why the software was modified away from
>playing good 'classical' chess to more powerful, deeper searching of
>tactical possibilities; this second option would give it a higher
>rating when measured against other machines but I think an improving
>beginner would find the 5 MHz Machine far easier to beat than the older
>Fidelity Challenger 9. Thus while it could be claimed that it is a
>stronger machine, it seems to be a much weaker player.

Having produced my own chess program I can make some educated guesses.

First, consider that the vast majority of chessplayers buying computer
chess machines are not that high rated a player-usually beginners.  The
higher a persons rating, the fewer there are and less of these would be
interested in fiddling with micro chess playing machines.

Second, the new weaknesses you pointed out would seem to all have a direct
bearing on move search time.  Lets take for example, pawn structure.  Just
to make sure that pawns are still linked after an exploratory move (not even
considering whether they're 'properly' linked for the given position) requires
a numerical evaluation factor be produced by examining for each pawn four (or
six) surrounding squares for other pawns of like color.  Take the bad bishop
detection.  Its easy for a human to look at the board and recognize a bad
bishop and its obvious when considering moves whether they impact such a
status. But a computer program must do costly scans accross the board to
determine such a thing.  And it would have to redo the scan thousands of
times during a move evaluation.  Now, I am simplifying things a bit since
there are programming tricks to shortcut some of this.  But the problem
remains that even the simplest 'extra' check added to a chess programs move
evaluation can have a severe impact on the number of nodes (moves) evaluated.

Third, the horizon effect critically impacts chess strength, especially at
lower ply levels that the micros are running at.  Therefore, there is the
constant conflict of "Will improving our program's chess 'knowledge'
overcome the weakness added due to shorter horizon forced on our search tree?"

Yet another reason to improve search depth is that the micro people would just
LOVE to beat BELLE or CRAY BLITZ, and they really don't have a ghost of a
chance unless their ply search depth as within a couple plies of those
monsters.  (In other words, no amount of chess smarts at 5 ply will beat a
dumb machine running 9 ply.  The 5 ply machine will start losing pieces
due to tactical miscalculations.)

Following the computer chess manufacterers like Fidelity over the years, one
can see how the various generations have gravitated first to one side and
then the other

Finally, I think you hit the nail on the head when you mentioned that the
new machine may well beat other machines even though its chess games may
not end up being as useful to a human learning chess.  There are several
entries in the marketplace, and with such fine margins of profit a dramatic
loss at a well recognized tournament against a competitor can well hit
the money belt for a year or more.

In summary, there are several nasty tradeoffs which must be made in deciding
the "personality" of a computer chess program.  I hope my "guesses" shed some
light on the subject.

Dave Trissel              {ihnp4,seismo,gatech}!ut-sally!oakhill!davet