Author: Charles Roberson
Date: 19:07:00 07/08/03
Hi Peter,
As you know, ANN's can be trained in mutiple ways and with many
architectures. As the KnightCap paper points out, multiple ANN's are good -- one
for each stage. TD learning has two methods: (1) learning from actual play and
(2) vicarious learning. There is a lot of promise in both, but I like the second
as it reduces run time. Also, using 3 layers, as opposed to two, improves the
learning rate. It is not necessary to use 3 layers becuase there is a proof that
2 layers are sufficient to learn any function.
Still, one might not use TD learning at all. I beleive training on a large
number of positions will work. The work by the DeepThought/Deep Blue team proves
this. Remember, a single layer ANN will produce the same results as the typical
best fit line effort. There are a few papers on the automated tuning by T.S.
Anantharaman in the ICCA journals and they use the best fit approach.
I beleive that graduated opponents or training data from graduated opponets
are necessary for the best learning. Many of the papers in the ICCA journal are
based on games or positions from GM vs GM. Max Euwe suggests, in his book
"Master vs Amateur", that amatuers couldn't learn from "Master vs Master"
because those types of positions didn't occur in their games against other
amateurs.
From this I plan to auto-tune my PE with an ANN using positions from GM vs
all levels of players. Now to define GM. I am not limiting my definition of GM
to humans.
I'll reiterate from my email -- one must train based on the final position
from each PV as opposed to the root position.
Charles
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.