On Apr 30, 2:34 pm, David Richerby
wrote:
Martin Brown wrote:
David Richerby wrote:

What do you mean by `percentage blunder rate'? The proportion of
the time that the GM plays a move that the engine thinks is, say,
more than one pawn worse than the best move?

That would probably do as a rough working definition. The search
depth or time might also need to be specified.

Sure.

And in fact the graph for %blunder rate for every player is in the
original article.

http://www.chessbase.com/newsdetail.asp?newsid=3455

That its shape broadly correlates with the rms error graph of the
players lends credence to the possibility that Crafty might have been
adequate for the task. And to be fair to the authors they did say that
others with access to the internals of stronger engines should repeat
their tests to see how they compare.

I would have liked to see the rms error graph with blunders excluded.
That might have shed some more light.

How does that make a difference?

Unforced tactical errors play their part in the outcome of games.
And these are precisely the sorts of thing that computer chess
engines are very good at spotting. Subtle long term structural games
are much harder for them to score.

So you're suggesting that ``Player X makes a one-pawn blunder in n% of
games'' is a better measure than ``Player X, on average scores n cp
lower per move.'' That does sound like a reasonable statement, though
I do worry that sacrifices of pawns are relatively common and might
still be mis-evaluated quite often. Kasparov used to sacrifice a pawn
for long-term initiative faster than you can say, ``My computer thinks
that's a pretty dodgy move.'' :-)

Although that may be true. If the program is analysing in blunder
check mode or classical analysis mode it will know the outcome of the
principle variation actually played as well as for its own
hypothetical better move(s).

Do you have any guess (or, shock!, data) on how often errors occur in
WC games that an engine (given reasonable time) would score down by
say 100cp?

It is in the paper referred to by this thread. Sadly the link to the
original article is broken. Anyone have a full copy?

Capablanca maintained a blunder rate of 0.01% (1 blunder in every
10000 moves) and the worst performer was Steinitz at 0.054% (blunder
every roughly every 2000 moves). These are interesting numbers and
right at the limits of human error rates for purely trivial mechanical
tasks like punch key data entry. It is quite astonishing how low these
are!

So as a rough guide if the average game lasts 40-50 moves (80-100
player actions) less than 4% of them will have their final outcome
determined by a blunder at GM level. So the other 96% of cases clearly
needs study.

It would be interesting to know if the downward march of error rate
with time in GM level play is actually due to improved training
methods or sparring against computers which always seize on any minor
tactical error. There look to be clusters of players from other eras
with similar error rates (and a few notable exceptions).

To put it into perspective commercial programming has an effective
error rate around 1-2% (and in some shops 10% is not unknown). Much
greater than 0.2% acheivable by the best formal development methods.
But competitive GM level chess is more than an order of magnitude more
accurate still.

Human error rates for various tasks are online at: http://panko.cba.hawaii.edu/HumanErr/

Regards,
Martin Brown