So change my function to
F[move] + G[Previous][move] + H[Penultimate][move] +
K[Penultimate] [Previous] [move]
Same space usage, but it has all the effects you are claiming and also O(1)
training time.
Here's the point: the Neural Network is not magic, and cannot detect
patterns faster than the underlying statistics allow.
From: [email protected]
[mailto:[email protected]] On Behalf Of Peter Drake
Sent: Wednesday, June 29, 2011 1:46 PM
To: [email protected]
Subject: Re: [Computer-go] A Linear Classifier Outperforms UCT on 9x9 Go
On Jun 29, 2011, at 10:17 AM, Brian Sheppard wrote:
Why is a classifier better than having a lookup table indexed by
OurLastMove, OppLastMove, ProposedNextMove that returns the Wins / Trials
experienced when ProposedNextMove is played after the sequence OurLastMove,
OppLastMove?
The advantage here is that we combine information from several piles:
- All times this move was played.
- All times this move was played in response to previous move X.
- All times this move was played in response to penultimate move Y.
The scheme you propose only gathers:
- All times this move was played in response to previous move X and
penultimate move Y.
This information is more accurate, but accumulates more slowly. (See the
Power of Forgetting paper for more discussion on this.)
Are the training cases for your classifier selected from only the UCT nodes,
or also from playout nodes?
>From the entire playout.
Is the output of your classifier used to initialize the Wins / Trials values
for legal moves in new UCT nodes? Is that done by assuming a fixed number of
trials (how many?) and setting Wins = ClassifierOutput * Trials?
There is no tree in this system. The primary policy (used for the first 10
moves of each playout) is to choose the (legal) move that the classifier
rates highest.
Is that the only use of the classifier in the system?
The above is the only use of the classifier.
Peter Drake
http://www.lclark.edu/~drake/
_______________________________________________
Computer-go mailing list
[email protected]
http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
