utility normalization and multiplication pi vs sigma

elefantopia

elefantopia

Member
1598411154684.png

current implementatin is a sigma: non normalized getutility are added
1598410646559.png
max scores: Amax+Bmax+Cmax+Dmax = ?
scores: 5+9+3+2 = 19 is not normalized and needs the value above passed somehow

infinite axis is a normalized pi : getutility are multiplied, response curve always between 0 and 1
max scores are always = 1 : Amax*Bmax*Cmax*Dmax = 1*1*1*1 = 1
so
score 0.5*0.3*0.1 is always normalized

benefit of the Pi :
  • no need to move the score max up the chain (which this implementation of Getutility doesn't even do)
  • no math explosion due to weights being too much somewhere down the chain
benefit of sigma:
  • if you want to boost a value you don't need to redo the other curves? maybe?
What's your thought on this
 
I might as well use your implementation this way it won't break my modification the next time you update BD.

1598465840566.png
The paper normalizes, how do you back propagate max score and what API do I need to call to normalize?
In the following scenario:
Utility Selector
| ... | ...
score1
|
...
|
scoreN
|
task
 
You could duplicate the Utility Selector and then make your own modifications to the GetUtility comparison. This is done in two separate methods so make sure you update them both.
 
You must log in or register to reply here.
Top