RE: computer-go: abstract info and neural nets

["Staelin, Carl" <staelin@xxxxxxxxxxxxxxx>]

Weimin,

Given the size of your networks, I think L-BFGS is
the way to go.  I think you will find that you get
much better convergence than with momentum backprop.
If you don't have some L-BFGS code of your own,
here is a pointer to some FORTRAN code:
	http://www.ece.nwu.edu/~nocedal/lbfgs.html
You should be able to use f2c to convert it to C,
assuming you are using C.

Cheers,

Carl

_________________________________________________
[(hp)]	Carl Staelin
	Senior Research Scientist
	Hewlett-Packard Laboratories
	Technion City
	Haifa, 32000
	ISRAEL
	+972(4)823-1237x221	+972(4)822-0407 fax
	staelin@xxxxxxxxxxxxxxxxx
_______http://www.hpl.hp.com/personal/Carl_Staelin_______


> -----Original Message-----
> From: Ran Xiao [mailto:ranxiao@xxxxxxxxxxxxxxxxx]
> Sent: Monday, January 14, 2002 12:01 AM
> To: computer-go@xxxxxxxxxxxxxxxxx
> Subject: Re: computer-go: abstract info and neural nets
> 
> 
> Carl,
> 
> Thank you for all the articles about the NN. They all are 
> very interesting
> and helpful. I hope one day I can build an artificial brain 
> likes Andrew of
> the movie "Bicentennial Man".
> 
> I am using quasi 3-layer NN (a 3-layer NN without sigmoid 
> function at the
> output layer). The number of input nodes is 81 for 9x9 board. 
> Right now, the
> number of hidden nodes is 160. The number of output node is 1 
> which is the
> plausible value [0, 1] for a given board for BLACK played.
> 
> For the board evaluation, my philosophy is that if the 
> players anticipated
> game correctly, the final score will be score for the first 
> move. Thus, if
> we have a correct full board evaluation function, no need to play GO.
> 
> The current training method thus is supervised 
> quasi-reinforced. I am using
> batch mode with back-propagation momentum-based weights 
> modification. I was
> thinking to use limited-memory BFGS.
> 
> Weimin
> 
> P.S. The following is a game I played with ForeverBlack for the latest
> update of NN weight file. The NN training is still far from 
> converge, so
> ForeverBlack behaves different every time I updated its 
> weight file. The
> play is very casual. ForeverBlack does not count stones, I 
> briefly counted
> it myself.
> 
> Date Sun Jan 13 13:07:16 PST 2002
> Boardsize 9
> Handicap  0
> Komi  5.5
> Black ForeverBlack
> White Human
> Score 11  43
>  B 0 D5
>  W 1 F5
>  B 2 E4
>  W 3 E6
>  B 4 A2
>  W 5 F4
>  B 6 F7
>  W 7 E7
>  B 8 B5
>  W 9 C7
>  B 10 A7
>  W 11 E3
>  B 12 C6
>  W 13 D6
>  B 14 D3
>  W 15 E2
>  B 16 B7
>  W 17 C8
>  B 18 H7
>  W 19 G6
>  B 20 F2
>  W 21 G2
>  B 22 G3
>  W 23 F3
>  B 24 C2
>  W 25 D2
>  B 26 D8
>  W 27 E8
>  B 28 G7
>  W 29 H6
>  B 30 E9
>  W 31 F8
>  B 32 D4
>  W 33 B8
>  B 34 H9
>  W 35 G8
>  B 36 B3
>  W 37 A8
>  B 38 H5
>  W 39 H4
>  B 40 G4
>  W 41 J5
>  B 42 E1
>  W 43 D1
>  B 44 B1
>  W 45 C1
>  B 46 G1
>  W 47 F1
>  B 48 C9
>  W 49 H3
>  B 50 D7
>  W 51 D9
>  B 52 D8
>  W 53 F9
>  B 54 D7
>  W 55 D9
>  B 56 D8
>  W 57 B9
>  B 58 J6
>  W 59 G5
>  B 60 B6
> 
> 
>