Re: [madman-discuss] how embarrassing...

[Shawn W. <sh...@wi...>]

=2D----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On Tuesday 30 December 2003 12:22 am, Andreas Kloeckner wrote:
> I get what you're saying, but only after reading through the
> description of your AutoDJ implementation.
>
> Of course, I had been brooding over how to best do the DJ for a while,
> and my idea went as follows:
>
> Assemble a list sorted by scores, throw out those that match a "never"
> criterion. Use a linearly (i.e. not uniformly) distributed random
> number=0D generator to pick from this. I'll illustrate that:

Interesting... two potential objections that jump out at me are:

=46irst, as you mentioned, the connection between score and play probabilit=
y=20
would be less clear.

Second, from an implementation perspective, having to sort the list might=20
be expensive with large collections.

On the other hand, as you also mentioned, my approach is extremely=20
sensitive to changes in low scores.  0 vs 1 should make a huge difference=20
(won't play to might play), but it's not so clear that 1 vs 2 should=20
double the play probability whereas 19 vs 20 increases it by just over 5%. =
=20
Also, with my approach, if you have a small number of high-score songs and=
=20
a large number of low-but-positive-score songs, you may not hear your=20
high-scoring songs much.  I'm not sure if that's a problem, or just=20
indicative that your scoring preferences need to be adjusted if that's not=
=20
what you wanted.

> If I understand your suggestion correctly, you mean to make the play
> probability depend linearly on the score, as in
>
>   probability ~ weighted_score * factor + offset
>
> I don't know whether you use "factor", but offset would be your notion
> of=0D "randomness" which would essentially level out the score differences
> between songs.

No, I don't have a scaling factor, and, yes, offset would be the notion of=
=20
"randomness", which really just levels out differences and boosts the=20
score of otherwise unmatched songs above 0, making them playable.

> I don't know how well either suggestion would do in real life. I'm
> looking forward to hearing your opinion on this.

I don't really know either.  I guess I'll find out how well my approach=20
works in practice.  If it turns out to be too hard to tune well, we'll=20
have to look at alternatives.

	Shawn.
=2D----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.2.3 (GNU/Linux)

iD8DBQE/8bXtp1Ep1JptinARAlxKAJ0ao5RpLrpUCZez8JdP8pTh/M5GFQCgh0Vr
lorCxvDyVkcwWwBuj1zQk7w=3D
=3D9BHe
=2D----END PGP SIGNATURE-----