Skip to content

Latest commit

 

History

History
46 lines (22 loc) · 1.53 KB

lecture-20.md

File metadata and controls

46 lines (22 loc) · 1.53 KB

Independent Component Analysis

Wikipedia Article on ICA

####Example: We have a group of people betting on a game. Our objective is to minimise the loss we make while betting.

  • There are multiple rounds of betting.
  • You must bet on every game.
  • One person in that group is an expert and is always correct.

Solution: Find the expert as quickly as possible.

xi(t) = prediction of the ith person in game t wi(t) = (boolean value) whether we listen to the ith person's advice z(t) = the winner.

Idea: Bet on the side which most people bet on. Only consider the bets of those people who have always been correct up to this point.

y(t) = [ Σi (wi(t) xi(t)) > 1/2 Σi wi(t) ]

Update the weight associated with each person (only those who have always been right before this game). If they guessed incorrectly in this game set their weight to 0 else 1.

wi(t+1) = wi(t) [ xi(t) = z(t)]

Regret : How much we will lose before we find the expert. Regret <= log2n

Update Problem: No one is always correct but some people are better then others at guessing.

wi(t) = weight associated with the ith persons advice.

Idea: Reduce the weight an individual has by 1/2 if they bet incorrectly. Keep their weight the same if they guess correctly.

wi(t+1) = wi(t) ( 1/2 [xi(t) = z(t)] + 1/2)