[MPlayer-dev-eng] comprehensive lavcopts test - in progress.

[Romain Dolbeau]

Roberto Ragusa wrote:

> May I suggest to use a genetic algorithm?
[snip nice explanations]
> This method is really powerful.

I'm trying that ATM. Seems pretty efficient. After exploring
0.001% of the design space, some options are already more
or less washed out from the top 102. Exemple:
"Float: [tcplx_mask]: null = 69 (67.6470588235294 %) tcplx_mask=0.1 = 28 (27.45098039215686 %) tcplx_mask=0.2 = 5 (4.901960784313726 %)"
That was after 340 results (out of 3359232 :-) were obtained
(initial step of 102 then 7 steps of 34), and the original
(random) distribution was:
"Float: [tcplx_mask]: null = 42 (41.1764705882353 %) tcplx_mask=0.1 = 30 (29.41176470588235 %) tcplx_mask=0.2 = 30 (29.41176470588235 %)"

If someone is interested (and has a lot of computing power :-),
I can supply the Java code. You need a recent JDK (on Linux/x86,
j2sdk1.4.2_01 works but j2sdk1.4.1_01 locks up sometimes).

Basically, the main process generate the options & handle
the genetic algorithm, and distribute the work through RMI
to RMI services on the server machines. Better have many
of those :-)

After each step, the code displays the best result (PSNR
wise), the fastest result within x% of the best result
(sometime losing 1% means a 10x speed increase :-), and
the above stats.

-- 
Romain Dolbeau