I tendto like exponentially weighted moving averages when I need a
fading memory. That keeps storage simple, updates fast, and nearly the
same effect
i.e.
wins = 0.99*wins + latest_result
sims = 0.99*sims + 1
Sent from my iPhone
On Jun 26, 2008, at 2:40 PM, "Ivan Dubois" <[EMAIL PROTECTED]> wrote:
This same topic already occured on the list some time ago.
I think the idea is to "forget" older results. For exemple you can
compute the win rate based only on the last 500 simulations. Older
information may not be up to date and will not help much because 500
simulations is enough to compute an accurate winrate.
The problem is that you have to store the result of 500 simulations
at each node. I think some people reported that it does indeed
increase the strength of their program.
----- Original Message ----- From: "Peter Drake" <[EMAIL PROTECTED]>
To: "Computer Go" <computer-go@computer-go.org>
Sent: Wednesday, June 25, 2008 5:48 PM
Subject: [computer-go] UCB/UCT and moving targets
UCB (and hence UCT) would treat the following sequences of wins
(1) and losses (0) the same:
01010101010101010101010101010101
00000000000000001111111111111111
11111111111111110000000000000000
Clearly, it would be better to favor the second sequence, because
that move has done more for us lately. Because the tree is growing,
the values of the moves are moving targets.
Has anyone done any work dealing with this phenomenon, e.g.,
somehow giving more weight to more recent playouts?
Peter Drake
http://www.lclark.edu/~drake/
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
_______________________________________________
computer-go mailing list
computer-go@computer-go.org
http://www.computer-go.org/mailman/listinfo/computer-go/
