- Calculatec=erfinv2(1 -ε), whereerfinv(x) is an inverse error function,εis a target classification error of the algorithm defined aserf(x) is the error function,h(ix) is a hypothesis formulated by thei-th weak learner,i= 1,...,M,αis the weight of the hypothesis.i
- Set initial prediction values:r1(x, y) = 0.
- Set "remaining timing":s1=c.
- Do fori=1,2,... untilsi+1≤0
- With each feature vector and its label of positive weight, associate
- Call the weak learner with the distribution defined by normalizingW(ix, y) to receive a hypothesish(ix)
- Solve the differential equationwith given boundary conditionst= 0 andα= 0 to findt=it* > 0 andα=iα* such that eitherγ≤νort* =s, whereiνis a given small constant needed to avoid degenerate cases
- Update the prediction values:r(i+1x, y) =r(ix, y)+αih(ix)y
- Update "remaining time":s=i+1s-iti