Talk:Projects/Plasma/Menu: Difference between revisions

From KDE TechBase
No edit summary
 
(Notes on Machine Learning for the menu)
Line 3: Line 3:
Why is the main goal is to be unique but not to be usable and convenient???
Why is the main goal is to be unique but not to be usable and convenient???
Regarding the ranking of program usage:
0) Knowing how long the program is in use for might not be important: I have kopete, kontact, amarok and klipper open constantly, but I only need to launch them from the menu when they crash. I would like to see what kind of results your program comes up with though, because apps like konq seem to be both always-open, and often getting new instances on different desktops, so it might work quite well.
1) If you're just pissed off with the "favourite applications" in kmenu (which is quite rubbish), take a look at quicklaunch. The dynamic mode on the quicklaunch kicker applet in kde3.5 is quite good at working out which applications I use most often(even takes note of what's launched with alt+f2). It isn't very consistant though(see point 2) so I still open quite a lot of stuff in konsole/alt+f2.
2) Changing the ordering of items in a list can *sometimes* make finding things take longer (because you can't use your spacial memory to *learn* where things are). It's probably not relevant in this case, but http://www.inference.phy.cam.ac.uk/dasher/ might be interesting to look at, as it makes important things promenant without changing the order.
3) Having a system that splits the day into 24 *independant* blocks might mean that the program takes a long time to learn. A hack that lots of people use in machine learning is to use blurring to make learning more general(ie add a gaussian (bell shaped curve) to the data, centered about the right time, rather than just a single value at the time you opened the program. That, or use a gaussian filter when *reading* the data). The width and height of the gaussian determine the amount of blur, so you can tweak it to your heart's content aftwerwards. Of course: if you're looking for real machine learning, you will probably get a long way with a little bayesian probability.
--alsuren

Revision as of 15:43, 28 May 2007

– The main goal of Raptor is to deliver a unique menu for KDE, which is easy … – Why is the main goal is to be unique but not to be usable and convenient???


Regarding the ranking of program usage: 0) Knowing how long the program is in use for might not be important: I have kopete, kontact, amarok and klipper open constantly, but I only need to launch them from the menu when they crash. I would like to see what kind of results your program comes up with though, because apps like konq seem to be both always-open, and often getting new instances on different desktops, so it might work quite well.

1) If you're just pissed off with the "favourite applications" in kmenu (which is quite rubbish), take a look at quicklaunch. The dynamic mode on the quicklaunch kicker applet in kde3.5 is quite good at working out which applications I use most often(even takes note of what's launched with alt+f2). It isn't very consistant though(see point 2) so I still open quite a lot of stuff in konsole/alt+f2.

2) Changing the ordering of items in a list can *sometimes* make finding things take longer (because you can't use your spacial memory to *learn* where things are). It's probably not relevant in this case, but http://www.inference.phy.cam.ac.uk/dasher/ might be interesting to look at, as it makes important things promenant without changing the order.

3) Having a system that splits the day into 24 *independant* blocks might mean that the program takes a long time to learn. A hack that lots of people use in machine learning is to use blurring to make learning more general(ie add a gaussian (bell shaped curve) to the data, centered about the right time, rather than just a single value at the time you opened the program. That, or use a gaussian filter when *reading* the data). The width and height of the gaussian determine the amount of blur, so you can tweak it to your heart's content aftwerwards. Of course: if you're looking for real machine learning, you will probably get a long way with a little bayesian probability.

--alsuren