To obtain an accurate model of a process the adaptation process should allow for an arbitrary accuracy within a given cost. Cost may be measured in terms of processing time or computing requirements. It is well known that to gain a better approximation of a process, the adaptation should be able to model a non-linearity at a desirable precision. Currently, methods that do so achieve their accuracy at a high computational cost. Furthermore, these methods do not guarantee i) optimal solution (neural networks), ii) convergence (extended Kalman filtering), or iii) manageable cost (Volterra systems). In this paper, we offer a simple yet powerful method, a switched filter bank, to this end.