|
Kiran
2015-04-26 13:56:47
|
|
I ran PBIL optimization through a couple of ranking systems (conceptually quite powerful), but the results are confusing, because the weights of the factors don't add up to 100%.
E.g. i have a system with 3 factors (F1, F2, F3)
a) the best fitness function yielded weights (0,0,1)
b) second best yielded (0,0,6)
1) Shouldn't the factors always add up to 100%
2) Not sure how (a) and (b) are different, because they both recommend using only F3.
3) Finally, if i have hierarchical nodes (e.g. F1->(F11, F12), F2->(F21,F22)), I noticed that the top-level hierarchy weights are not optimized - it assumes equal weightage.
This could give sub-optimal results, and it seems it's best to keep things simple and have 1 flat list of all ranking factors used in optimization.
4) Is Genetic Optimization relevant for Ranking?
Any guideline document/blog outlining how to best use PBIL and/or Genetic optimization of ranking systems and interpret results would be useful.
thanks
Kiran
|
|