[gnugo-devel] predictability vs. strength

[Douglas Ridgway]

There's a tradeoff between predictability and strength. As a way of
exploring this, in a quick and dirty way, I replaced the 0.01 constant
random multiplier in value_moves.c with various other values and looked at
sets of 100 9x9 games against the unmodified code (5.5 komi, alternating
colors, 3.5.6+, level 10). Of interest is the win percentage and the
number of unique games.

Eps     p_win/%                 unique
0.01    50 (by definition)      88
1.0     45                      93
10.0    34 (*)                  99

Only at 10.0 do the repeats almost disappear (working hard to play 100 
unique games out of about 10^100 available), by which point the weakness 
is obvious (and statistically significant). Being less predictable is not 
so easy, there are not lots of nearly as good moves sitting just below the 
best move.

I also tried a set of 100 of the standard code with --nofusekidb on both 
sides. There were 8 unique games, indicating that most of the variation 
GnuGo has in self play comes from the fuseki db, at least on 9x9.

doug.