最小均方误差推导(RPML )

最小均方误差推导(RPML )
最小均方误差推导(RPML )推导如下


A=tΦwA=t-\Phi w
dA=tΦdw=ΦdwdA=t-\Phi dw=-\Phi dw
f=ATRAf=A^TRA
df=(dA)TRA+ATRdA df=(dA)^TRA+A^TRdA df=tr((dA)TRA+ATRdA) df=tr((dA)^TRA+A^TRdA) df=tr(ATRTdA+ATRdA) df=tr(A^TR^TdA+A^TRdA) =tr(AT(RT+R)dA)=tr(A^T(R^T+R)dA) =tr(2ATRdA)=tr(2A^TRdA) df=tr(2ATRΦdw)df=-tr(2A^TR\Phi dw)
由于df=(fw)Tdwdf=(( \frac{\partial f}{\partial w} )^Tdw )
fw=ΦTRTA\frac{\partial f}{\partial w}=\Phi^TR^TA=ΦTRT(tΦw)=\Phi^TR^T(t-\Phi w)=ΦTR(tΦw)=\Phi^TR(t-\Phi w)
fw==0\frac{\partial f}{\partial w}==0ΦTRt=ΦTRΦw\Phi ^TRt=\Phi^TR\Phi ww=(ΦTRΦ)1ΦTRtw^{*}=(\Phi^TR\Phi)^{-1}\Phi^T Rt