Spline procedures have proven effective in estimating smooth functions. How
ever, spline procedures based on stepwise addition and/or deletion have som
e drawbacks. They suffer from the knot compounding problem, making their pe
rformance suboptimal. Furthermore, due to computational complexity, spline
procedures may not achieve their full potential. In this article, we propos
e a novel knot selection algorithm for regression spline estimation in nonp
arametric regression. The algorithm includes three new components: knot rel
ocation, guided search, and local fitting. The local properties of the spli
ne functions are used to efficiently implement the algorithm. Extensive sim
ulation studies are performed to demonstrate the improvement of the new kno
t selection algorithm over the stepwise addition and deletion scheme, and t
he advantages of the spline procedure with the new knot selection scheme ov
er alternative adaptive methods. In the simulations, our procedure achieves
very competitive performance with alternative methods and has substantial
advantage in nonsmooth functions. Finally, the usefulness of the proposed m
ethod is illustrated by an application to signal recovery in speech signal
processing.