Tony,

I'd like to start off by saying two things: that the whole point to the learning routine is to just improve over a regular shuffle. Not to be the pancea of all time and space. Also, squash will -eventually- have all those tweaking playlist thingamabobs. But there is little point in doing that until the baseline works (i.e. "adv." shuffle, and rest of the interface and guts).

Ok, so anyway. To address your concerns: Statistics already can be gathered for multiple people. But the only interface currently for switching people is to restart the player.

Concerning skipping a song you like or listening to a song you hate. That's why it's called statistics? You are likely to listen to the former more often than the latter, but not garuanteed to always listen to or to always skip. The more you like or dislike the more extreme the rating. I mean, isn't this all a definition how statistics work? There are people who have no children or 10 children even if the average is 2.5. That's just life.

As for manually scoring. I don't mind implementing this, but just like custormizing the playlist, this is alot of work for the user. (So in other words, I'd rather not bother with it personally -- though I understand others wanting to). So I don't have that feature in just yet, but don't worry it will go in eventually (feel free to contribute ).

Concerning decaying averages. You have a good point. However automatic data collection is -so slow- already. It has been 3 months and I have only listened to 70% of all my music once (of course 5% I skipped). So maybe in another month I'll be start round two of skip vs play. Decay for this kind of input rate would need to be very small or else you will lose all of you data. I will not be implementing decay soon, since it could easily destroy the whole system if done badly -- but I suppose it could be added safely if you are careful (I, as a user, will just turn it off though).

Again I will eventually have playlist manual manipulation, but that's like version 1 going to version 2 features. Not the version 0 to version 1 features.

Lastly remember that squash effectively works like an improved shuffle -- so very good songs just appear more frequently than good songs. And bad songs can appear too, but less frequently. See another post I'm about to right for some numbers. The important part is that it's shuffle of all the music, not some magic dijinn that knows exactly what you want to listen to right now.

Statistics can only play the songs you like better more frequently and vice versa. And the level of data that is being collected right now has no way of knowing what you are -in particular- would like to hear. Just what you in general like more than otherwise.

Thanks for your comments I like discussing this. And I plan on making it my master's thesis, so this helps me later in defending claims.

p.s. Thinking about the decaying average more. This seems like overkill, even if the data was collected at a quicker rate somehow. The current goodness function is: (p-s)/(p+s+1). Which "decays" if say you played the song the first 10 times, but then decide you hate it. Anyway, it may still be worthwhile -- but I think it needs to be very small.