Description
In a $J$ class classification problem with data of the form $(j_n,\underline(xn)$, $n=1,\ldots,N$ where $j_n \in \(1,\ldots,J\)$ and $\underline(x)_n = (x_(1n),\ldots,x_(Mn))$, linear discriminant analysis produces discriminant functions linear in $x_1,\ldots,x_M$. We study a procedure which contructs discriminant functions of the form $\sum_m\varphi_m(x_m)$, where the $\varphi_m$ are nonparametric functions derived from an iterative smoothing technique. Judging from a variety of data sets, the method offers promise of being a significant improvement on linear discrimination.