Machine Learning Experience

[Mike Valvo]

For what it is worth, I would like to describe an experience with
a learning algorithym built into BEBE a couple of years ago. I am
not fully aware of how it worked, but it went something like this:
The machine would use its evaluations of positions and if the
evaluation fell, it would attempt to improve it the next time the
position came up. The key was the transpositional table: it was
not turned off between games!

We played the Marshall Counter Attack in the Ruy Lopez every game.
Initially the machine played poorly (it had the book turned off),
but it improved its play each game. The was a tremendous benmefit
for me: I got to learn through experience why certain things were
good or bad.

It wasn't necessary for me to even play the best moves for the
machine to improve its play in a subsequent game. The machine
will be looking at all my possibilities even if I don't play
them myself. Of course, if I play the best, the machine's
algorithym will work better.

--Mike Valvo