A composite system design methodology for instrumentation and embedded system

Surely, the main goal when designing an embedded system is performance maximization. Nevertheless, physical constraints such as silicon area and/or power consumption take and active part in the system design by limiting, most of time, the solution space and hence reducing the system performance. In this paper we present a methodology for selecting the best composite model for an embedded system in a constrained environment. The envisaged constraints are computational complexity and latency which respectively address the computational complexity issue in SW and HW realizations, respectively. It is assumed that the algorithm to be implemented in the embedded system is not given and must be constructed by relying on some (input, output) examples. Models considered for the system identification phase can be linear (e.g., ARMAX), nonlinear (e.g., neural based models) or composite (a suitable mix of linear and nonlinear models). The best solution is then selected from the candidate ones to optimally satisfy the application requirements