1 Introduction

The recent launch of the SOlar and Heliospheric Observatory (SOHO) satellite, has renewed interest in the classification (Seely et al. 1997; Laming et al. 1997) and interpretation (Brekke et al. 1997; Judge et al. 1997) of high spectral resolution ultraviolet (UV) and extreme ultraviolet (EUV) emission spectra. The majority of these spectra come from the Solar Ultraviolet Measurement of Emitted Radiation (SUMER), and Coronal Diagnostic Spectrometer (CDS) instruments onboard SOHO (Wilhelm et al. 1995; Harrison et al. 1995).

A first step in the analysis of emission line spectra is to identify and measure properties of lines believed to be present. This is usually achieved by associating (subjectively) the observed spectral profiles with ionic and atomic transitions of "known" laboratory wavelengths. From these possibly biased decompositions, physical models of the underlying plasma are sought. In an effort to obtain the best possible scientific results from their spectra, the CDS and SUMER teams have set about ways to produce the most "reliable" decomposition; see Brynildsen (1994) for more details.

Standard spectral decomposition techniques unfortunately prove to be unstable when presented with data of low signal to noise ratio, or data that is poorly sampled. In particular these instabilities cause subtle differences in the decomposition of each spectrum and can lead to significantly different physical interpretations. This has prompted us to search for a method that can provide spectroscopists with reliable decompositions of observed spectra that are as free as possible from subjective bias.

We use a heuristic approach to decomposition. We use a Genetic Algortihm (GA) to fit model line profiles, which for our purpose we chose to have Gaussian form, to provide a simple parameterisation of the spectrum under analysis. This approach exploits the stability and optimization capabilities of natural selection (Darwin 1859). Sections 2.1 and 2.2 describe the basic GA formalism, and an introduction to our Gaussian fitting GA, hereafter Ga-GA.

The GA technique is applied under ideal conditions (to "simple" noiseless test spectra) in Sect. 3.1. This first test also helps to highlight how well genetic operators are suited to this task. Section 3.2 gives a much more stringent test of the how a GA performs when fitting spectra containing unstructured random noise. Here, the GA's stability in the presence of random Gaussian noise is compared to that of standard profile fitting and optimization algorithms. We show that these standard algorithms are blighted by possible user bias which is not present with the GA technique. To aid further comparison of our GA technique to standard analysis algorithms we have constructed model spectra with realistic noise and continuum/background levels. The results are discussed in Sect. 3.3.

The ability of the GA approach is given a final test in Sect. 4 on quiet Sun SUMER spectra. There we compare our results with those obtained from an analytical decomposition performed by Judge et al. (1997). We note that their technique used additional information not available to the GA.

Although much emphasis must be placed on the fact a GA requires minimal user input, in certain circumstances user input can prove useful, such as cases where relative wavelengths and intensities are well known from atomic physics. Such additional constraints can (almost trivially) be "hard-wired" into the algorithm. Section 4.1 highlights the possibilities of applying rigid a priori constraints to the observed spectrum.