Kevin Pacey wrote on Tue, Mar 15, 2016 03:09 AM UTC:
On a Canadian chess message board (chesstalk) a poster (Mario
Moran-Venegas) wrote today, re: AlphaGo's loss in game 4 of the match:
"AlphaGo's Policy neural network does not care about the quantity of point
lost or won by. It's highest priority is maximizing winning probability or
(when losing) minimizing losing probability. After move 78, it should have
followed what you say chess engines do: attempt to prolong the game by
complicating it.
The Policy neural network is the boss of hundreds (literature says a max of
1200, I don't know how many were actually used) of brute force engines
similar to chess engines.In the future versions I see the following
changes:
1.A change in the dynamic depth-of-analysis assigned to an engine. My guess
is that which ever engine was given the task of tackling the area around
move 78 was not going deep enough thus affecting the overall assessment of
the entire board as a loss for Black (AlphaGo). Many (including commentator
Michael Redmond) are now saying move 78 "did not work".
2.A change on the Policy NN to make use of complexity on the board."
If Mario's guess above, concerning move 78 of game 4, is correct, an
implication might be that the large 19x19 board used for Go may be close to
the upper limit of what the neural net technique (plus brute force engines)
used is currently capable of allowing a computer to beat top humans at, for
the game of Go, as played on an nxn size board. Make n significantly larger
than 19, that is, and a computer might fail to beat the top human players.
Not sure if the same would apply for a very large board size chess variant
too, as more calculation than intuition would be used than for Go, but
maybe there's something to the idea.
Much earlier in this thread Joe Joyce mentioned a very large board war game
of his invention that was arguably a chess variant, too. Fwiw, I've
invented a couple of 5x5x25 4D chess variants (625 cells) which would have
more cells than standard 19x19 Go, though I recall Joe's war game was even
larger. In any case, a very large board size might appeal to more potential
players of a given chess variant than other ideas that have been mentioned
in this thread, such as changing the rules a game is played by every other
turn (or the same for how a given piece moves).
