Object-oriented model for partially separable functions in parameter estimation
In parameter estimation, a model function depending on adjustable parameters is fitted to a set of observed data. The parameter estimation task is an optimisation problem, which needs a computational kernel for evaluating the model function values and derivatives. This article presents an object-ori...
Saved in:
| Main Author: | |
|---|---|
| Corporate Author: | |
| Format: | Article |
| Published: |
1999
|
| Series: | Acta cybernetica
14 No. 2 |
| Kulcsszavak: | Számítástechnika, Kibernetika |
| Subjects: | |
| Online Access: | http://acta.bibl.u-szeged.hu/12627 |
| Summary: | In parameter estimation, a model function depending on adjustable parameters is fitted to a set of observed data. The parameter estimation task is an optimisation problem, which needs a computational kernel for evaluating the model function values and derivatives. This article presents an object-oriented framework for representing model functions, which are partially separable, or structural. Such functions are commonly encountered, e.g., in spectroscopy. The model is general, being able to cover a range of varying model functions. It offers flexibility at runtime allowing the construction of the model functions from predefined component functions. The mathematical expressions are encapsulated and a close mapping between mathematics and program code is preserved. Also, all interfacing code can be written independently of the particular mathematical formula. These properties together make it easy to adapt the model to different problem domains: only tightly controlled changes to the program code are required. The paper shows how derivatives of the model function can be computed using automatic differentiation relieving the programmer from writing explicit analytical derivative codes. The persistence of the objects involved is discussed and finally the computational efficiency of the function and derivative evaluation is addressed. It is shown that the benefits of the object-oriented model, namely the higher abstraction level and increased.flexibility, are achieved with a very moderate loss of performance. This is demonstrated by comparing the performance with low-level tailored C-code. |
|---|---|
| Physical Description: | 285-302 |
| ISSN: | 0324-721X |