Edited by author 10-23-2001 06:42 AM
After reading Freund and Schapire, the first question that comes to mind is: if a classifier does worse than random guessing, why not just negate it? Now its better ;-)

  The other thing that came to mind is that if you have some Bernouli trial (a classification) you can take a series of them and you have a binomial distribution. I can't remeber too much of the math right now, but as I recall there's some way to extract a much better classification from the binomial than from the individual Bernouli trials. So myquestion is then how do these boosting algorithms compare to simple combination of single classifications. Does one mathematically reduce to the other? To get a binomial I believe that the Bernouli trials must be independent- how independent are the different sets of training examples that Freund and Schapire use?